Squash history (masters/, web/; runs/ replaced by runs v2)

.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,37 @@
+---
+license: other
+pretty_name: Embodied First Exam media
+tags:
+- robotics
+- embodied-ai
+- code-agents
+- simulation
+---
+
+# Embodied First Exam: media
+
+This repository holds the pictures, videos and agent runs behind the [Embodied First Exam](https://www.embodied-first-exam.ai), a benchmark for code agents in robotics. Each question in the exam is a frozen simulation scene with a goal written for a person, and the answer is a program, which is marked by replaying it in a fresh simulator.
+
+## Layout
+
+```
+masters/<suite>/<task>/video.mp4     master video of the task being solved (2560x1440 or larger, ray or path traced)
+masters/<suite>/<task>/poster.png    a full-resolution frame
+masters/<suite>/<task>/meta.json     how it was made: source run, renderer, size, speed, success check
+masters/<suite>/teaser.mp4           the suite's teaser, and highlights.json
+web/<suite>/...                      the compressed copies the website plays
+runs/index.json                      agent runs that can be watched step by step
+runs/<run>.json, runs/<run>/         one run: every step, tool call and picture the agent saw, and the verifier's replay
+```
+
+## How the masters are made
+
+Since 16 October 2026 every master is rendered natively above 1080p (2560x1440 is the target) with ray tracing or path tracing; rasterized renders are not accepted. Each video replays a solved run of the graded task itself (the reference solution, a successful agent run, or an official demonstration), with physics untouched; only the camera, lights and materials change. Tasks without a solved run show a slow orbit of their own starting scene. The full standard is `masters/MEDIA_STANDARD.md`.
+
+## Contributing
+
+Suite authors send their media through a review repository; after review it is mirrored here. Please do not open pull requests on this repository.
+
+## Licence
+
+The videos and pictures are renders of the suites' simulation scenes, made by the suite authors and the exam's maintainers. The scenes and assets keep the licences of their upstream benchmarks; see each suite's page on the exam's website.

masters/.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

masters/MEDIA_STANDARD.md

@@ -0,0 +1,223 @@
+# High-quality media for exam suites (v2, 2026-10-16)
+
+> **v2 (2026-10-16)**: masters must be above 1080p (2560x1440 is the target), and every frame must be ray traced or path traced. Rasterized renders are no longer accepted. See section 9. v1 was 2026-10-07.
+
+The Embodied First Exam site shows two kinds of media from every suite:
+
+- one **teaser video**, which plays in the scrolling wall on the home page and on the suite page;
+- one **media item per task**, which appears in the task list.
+
+Most of what we show today is far below the quality of the simulators we use: 360x360 clips, 480p screen grabs at 15 fps, and one generic scene image shared by dozens of tasks. This page sets the bar and explains how to reach it.
+
+![RoboLab, before and after](standard_examples/robolab_before_after.jpg)
+
+![BEHAVIOR-1K, before and after](standard_examples/behavior_before_after.jpg)
+
+## 1. The bar
+
+1. **Resolution: above 1080p.** Render natively at 2560x1440 (2K), 16:9. 1920x1080 is no longer accepted. 3840x2160 counts as excellent.
+2. **Ray tracing or path tracing, always. Never rasterization.**
+   - Rasterizers are not accepted. These include OpenGL/EGL renders, MuJoCo's built-in renderer, Genesis's default rasterizer, SAPIEN's default shader, PyBullet's renderer, and Blender Eevee or Workbench.
+   - Section 2 lists the switch for each simulator.
+   - If your simulator has no ray tracer, record the poses per frame and render them in Blender Cycles (section 2). A cinematic look is welcome there.
+3. **A hard teaser.** Pick one of your suite's hardest tasks, ideally one that no evaluated agent solves (for example, GPT-6 Astra fails it), and show it solved.
+4. **Every task gets media.** A video is the default, made to the same spec. Use a still only when a video is truly impossible, and the still must still be high quality and show the task clearly.
+
+| | Required | Excellent |
+|---|---|---|
+| Resolution | above 1920x1080, 16:9, rendered at that size (no upscaling); the target is 2560x1440 | 3840x2160 |
+| Frame rate | 30 fps, constant | 60 fps |
+| Rendering | ray tracing or path tracing (never rasterization), full-resolution textures | path tracing with a denoiser, e.g. Blender Cycles with a cinematic look |
+| Encoding | H.264 or H.265, yuv420p, CRF 18 or lower | CRF 16, or lossless frames |
+
+Deliver master files. The site makes its own web copies, so never compress for the web yourself.
+
+## 2. Recipes by simulator
+
+All of these were used for the 2026-10-07 renders of BEHAVIOR-1K, RoboLab, ManiGuard and RoboWits. Timings were measured on a shared RTX PRO 6000.
+
+### Isaac Sim / Isaac Lab (e.g. RoboLab)
+
+Switch the renderer to path tracing:
+
+```python
+import carb
+s = carb.settings.get_settings()
+s.set("/rtx/rendermode", "PathTracing")
+s.set("/rtx/pathtracing/spp", 16)            # samples per pixel per frame
+s.set("/rtx/pathtracing/totalSpp", 16)
+s.set("/rtx/pathtracing/maxBounces", 6)
+s.set("/rtx/pathtracing/optixDenoiser/enabled", True)
+s.set("/rtx/pathtracing/fireflyFilter/enabled", True)
+```
+
+- After each physics step, render enough sub-frames to accumulate the samples before you capture the frame. With Replicator, that is `rt_subframes`.
+- Render only the presentation camera. Shrink or switch off the observation cameras.
+- Timings:
+  - 1080p at 16 spp: about 0.45 s per frame;
+  - 1440p at 32 spp: about 0.95 s per frame.
+- Reusing one Isaac app for many tasks cuts scene setup from about 140 s to a few seconds.
+
+### OmniGibson (BEHAVIOR-1K, ManiGuard)
+
+- Path tracing works the same way as in Isaac above.
+  - RTX Real-Time 2.0 (`RealTimePathTracing` in recent Isaac Sim) is a good fast option: about 0.3 s per frame at 1080p.
+  - Full path tracing at 8 spp plus OptiX takes about 1.4 s per frame at 1440p.
+- **Recorded demos with full states** (for example the official BEHAVIOR Challenge raw data):
+  - restore the state per frame with OmniGibson's `DataPlaybackWrapper`;
+  - no physics runs between captured frames, so a time-lapse renders only every k-th frame, and the playback is exact.
+- **House-scale scenes:**
+  - use a chase camera with smoothing of about 0.5-1.5 s on position and yaw, aiming at the torso and hands;
+  - check line of sight against walls, so the robot never disappears behind one.
+- Switch off the robot's own cameras and the viewer camera while rendering.
+
+### Genesis (e.g. RoboWits)
+
+- **Use Nyx**, the path tracer of Genesis World 1.x:
+  - packages: `pip install genesis-world gs-nyx`;
+  - add it to the scene as a camera sensor: `scene.add_sensor(NyxCameraOptions(res=(2560, 1440), ...))`;
+  - we rendered with FastPathTracer at 128 spp, the OIDN denoiser, and a Poly Haven HDRI (CC0) plus one key light;
+  - 0.1-0.2 s per frame at 1440p.
+- **If your suite pins an older Genesis for physics:**
+  1. replay the run in the pinned version and record every link's pose per frame;
+  2. rebuild the scene in Genesis World 1.x and set those poses frame by frame.
+  - Check that the rebuilt links match the replay. Ours matched to within 1 micrometre.
+- Genesis's default camera is a rasterizer and is not accepted. Use Nyx.
+- Draco-compressed `.glb` meshes crashed Nyx. Re-export them uncompressed.
+- Genesis CPU physics can diverge between CPU vendors. Replay on the same kind of machine that recorded the run, and check success.
+
+### SAPIEN / ManiSkill
+
+Use the ray-tracing shader packs on the render camera:
+
+```python
+env = gym.make(env_id, render_mode="rgb_array",
+               human_render_camera_configs=dict(width=2560, height=1440, shader_pack="rt"))
+```
+
+`rt-med` and `rt-fast` trade quality for speed. The default shader is a rasterizer and is not accepted.
+
+### MuJoCo, robosuite, PyBullet and others
+
+MuJoCo's built-in renderer, robosuite's and PyBullet's are rasterizers, so they are not accepted at any resolution. Render in Blender Cycles instead:
+
+1. Replay the run and record every body's pose per frame (or the recorded states).
+2. Import the same meshes into Blender and set those poses frame by frame. No physics runs in Blender.
+3. Render with **Cycles** on the GPU. **Eevee and Workbench are rasterizers and are not accepted.**
+
+A cinematic look is welcome. This is what the RoboARC re-render of 2026-10-14 used: Blender 4.5 LTS, about 1.4-6 s per frame at 1440p on an RTX 4090.
+- **Sampling:** Cycles with OptiX, adaptive sampling (threshold 0.01, at least 64 samples), and the OIDN denoiser with albedo and normal passes.
+- **Set and light:** a seamless studio floor, one large soft key light, a weak overhead softbox and a rim light, and a dim sky dome.
+- **Colour:** AgX view transform, Medium High Contrast look.
+- **Camera:** a slow, eased orbit of about 10 degrees.
+- **Optional:** a gentle depth of field and motion blur, as long as the robot and the goal objects stay sharp.
+
+## 3. Camera and picture
+
+- **Use a dedicated render camera.** Never use the agent's observation camera, a viewer window or a screen recording.
+- **Framing:**
+  - a 3/4 elevated view, fixed or moving slowly;
+  - the robot and every object the goal mentions stay visible;
+  - the objects are large enough to read.
+- **Keep the key action inside the central 80% of the frame.** The site crops tiles to other shapes.
+- **Clean picture:** no text, logos, watermarks, UI, debug markers or axes. Balanced exposure: no blown-out windows, no crushed blacks.
+- **One look per suite:** the same camera style and lighting across all of its tasks.
+- **Poster:** a frame that reads on its own, usually the solved state with the objects in view. Not a close-up of the robot's back.
+- **Length:**
+  - a teaser runs 10-40 s, a task video 5-60 s;
+  - long tasks may be sped up, as long as the speed stays constant and is recorded in `meta.json`.
+
+## 4. Which run to show
+
+- **Show the graded task itself**: the same scene, instance and goal.
+- **Source**, in this order:
+  1. the reference solution (oracle);
+  2. a successful agent trial;
+  3. an official demonstration of the same task.
+- **Do not change physics.**
+  - Re-simulate the recorded actions exactly as the verifier does, or play back recorded states. Only the camera and the render settings may change.
+  - If your simulator is not deterministic, check the task's success condition at the end of the re-render.
+  - **Never publish a failed replay as a solved run.**
+- **No solved run?** Render a 6-8 s slow orbit of that task's own initial scene, plus a poster. Never reuse a generic scene image.
+
+## 5. What to deliver
+
+Upload everything to the private Hugging Face dataset [`the review repository`](the review repository). Its README has the commands. Uploads arrive as pull requests that a maintainer merges.
+
+For each task, in `<suite>/<task_id>/`:
+
+| File | Content |
+|---|---|
+| `video.mp4` | the master video, to the spec in section 1 |
+| `poster.png` | one full-resolution frame, chosen as in section 3 |
+| `still.png` | only for tasks without a video: above 1920x1080 (2560x1440 target), ray or path traced, showing the robot, every goal object and the goal (for example, start and goal side by side) |
+| `meta.json` | see below |
+
+For the suite: `<suite>/teaser.mp4`, plus its poster.
+
+```json
+{"suite": "robolab", "task": "put_two_mugs_on_shelf", "source": "agent-success:<run>/<trial>",
+ "renderer": "isaac-rtx PathTracing spp=32 optix", "rendering": "path-tracing", "width": 2560, "height": 1440, "fps": 30,
+ "speed": 8.25, "success_check": true, "image": "rcb-robolab:0.1.1"}
+```
+
+`source` is one of `oracle`, `agent-success:<run>/<trial>`, `official-demo:<dataset>/<episode>` or `scene-orbit`.
+`rendering` is `path-tracing` or `ray-tracing`. `renderer` names the renderer and its settings.
+
+## 6. Check before you hand in
+
+Run `python3 check_media.py <suite_dir>`. It needs ffmpeg.
+
+- **It fails on:**
+  - a resolution of 1920x1080 or lower, a wrong aspect, or a frame rate below the bar;
+  - a `rendering` other than `path-tracing` or `ray-tracing`, or a `renderer` that names a rasterizer (OpenGL, EGL, Eevee, Workbench, ...);
+  - a codec other than H.264/H.265;
+  - an encode above CRF 18;
+  - missing files or `meta.json` fields;
+  - a failed replay marked as a solved run.
+- **It warns about:** a size below the 2560x1440 target, a variable frame rate, black bars, and lengths outside the ranges above.
+- **A maintainer then looks at every teaser** and at a sample of task videos before anything goes on the site.
+
+## 8. Addendum 1 (2026-10-08): every task in motion, and your five hardest
+
+1. **A video for every task.** Replay every task and record the robot doing it: the reference solution, a successful agent run, or an official demonstration.
+   - Use a still, or an orbit of the starting scene, only for a task with no solved run at all.
+   - Say why in that task's `meta.json`, as `"why_no_video"`.
+2. **The robot moves.** In a teaser, and in every video meant for the home page, the robot itself must move: an arm reaching, a hand grasping, a body walking. A moving camera over a still scene does not count.
+3. **Your hardest five.** Pick at least 5 tasks that are hard and interesting, each with a video where the robot moves. List them in `<suite>/highlights.json`, best first:
+
+   ```json
+   {"teaser": "<task_id>", "highlights": ["<task_id>", "<task_id>", "<task_id>", "<task_id>", "<task_id>"], "why_fewer": ""}
+   ```
+
+   - Hard means that agents fail it, or that it takes real skill: a long horizon, precision, a tool, two arms, balance.
+   - The teaser is one of the highlights.
+   - Fewer than 5? Give the reason in `why_fewer`, for example "the suite has only 4 tasks" or "only 3 tasks have a solved run".
+4. **Why it matters.** The home page's wall gives every suite the same number of tiles and fills them from `highlights.json`. Each suite's cover is its teaser. A suite without highlights gets fewer tiles.
+
+`check_media.py` now fails a suite in four cases:
+- it has no `highlights.json`;
+- a highlight has no video where the robot moves;
+- it lists fewer than 5 highlights without a `why_fewer`;
+- a still or orbit has no `why_no_video`.
+
+## 9. Addendum 2 (2026-10-16): above 1080p, ray or path tracing only
+
+1. **Above 1080p.** Masters must be larger than 1920x1080; render natively at 2560x1440 (2K) or at 4K. Posters and stills follow the same rule.
+2. **Ray tracing or path tracing only.** Never rasterization: no OpenGL/EGL renders, no MuJoCo, robosuite or PyBullet built-in renderers, no Genesis default camera, no SAPIEN default shader, no Blender Eevee or Workbench.
+   - If the simulator has no ray tracer, render the recorded poses in Blender Cycles. A cinematic look is welcome; section 2 has a recipe.
+3. **`meta.json` gains `"rendering"`**: `"path-tracing"` or `"ray-tracing"`.
+4. **Scope.** This applies to every upload from 2026-10-16 on, re-uploads included. Media already on the site stays until its suite re-uploads.
+
+`check_media.py` now also fails on:
+- a video, poster or still of 1920x1080 or less;
+- a missing or unknown `rendering`;
+- a `renderer` that names a rasterizer.
+
+The checks in sections 6 and 8 still apply.
+
+## 7. Examples
+
+![RoboLab, path traced at 1440p](standard_examples/robolab_example.jpg)
+
+![ManiGuard, oracle replay with zero safety violations](standard_examples/maniguard_example.jpg)

masters/MEDIA_STANDARD.zh.md

@@ -0,0 +1,222 @@
+# 套件媒体的高质量标准（v2，2026-10-16）
+
+> **v2（2026-10-16）**：母版必须高于 1080p，目标是 2560×1440（2K）；每一帧都必须用光线追踪或路径追踪渲染，不再接受光栅化。详见第 9 节。v1 发布于 2026-10-07。
+
+Embodied First Exam 网站会展示每个套件的两类媒体：
+
+- 一段 **teaser 视频**：在首页的滚动墙和套件页里播放；
+- 任务列表里**每道题各一个媒体**。
+
+现在网站上的大部分素材，远远配不上我们用的仿真器能达到的画质：有 360×360 的小视频，有 15 fps 的 480p 截屏，还有几十道题共用一张通用场景图。这份文档定下标准，并说明怎么做到。
+
+![RoboLab 前后对比](standard_examples/robolab_before_after.jpg)
+
+![BEHAVIOR-1K 前后对比](standard_examples/behavior_before_after.jpg)
+
+## 1. 标准
+
+1. **分辨率必须高于 1080p。** 直接按 2560×1440（2K）渲染，画幅 16:9。1920×1080 不再接受。3840×2160 算优秀。
+2. **必须用光线追踪或路径追踪，绝不能用光栅化。**
+   - 光栅化渲染一律不收，包括 OpenGL/EGL 渲染、MuJoCo 自带渲染器、Genesis 默认光栅化相机、SAPIEN 默认 shader、PyBullet 渲染器，以及 Blender 的 Eevee 和 Workbench。
+   - 各仿真器怎么开见第 2 节。
+   - 仿真器没有光追的，逐帧记录位姿，再到 Blender Cycles 里渲染（见第 2 节）。在 Blender 里欢迎做电影感的效果。
+3. **teaser 要选难题。** 从套件里挑最难的题，最好是参评 agent 都做不出来的（比如 GPT-6 Astra 失败的），并展示它被做成的过程。
+4. **每道题都要有媒体。** 默认是视频，规格同上。只有实在做不出视频时才用静图，静图也必须高清，并且能看清这道题要做什么。
+
+| | 必须 | 优秀 |
+|---|---|---|
+| 分辨率 | 高于 1920×1080，16:9，按这个尺寸直接渲染（不许放大）；目标 2560×1440 | 3840×2160 |
+| 帧率 | 30 fps，恒定 | 60 fps |
+| 渲染 | 光线追踪或路径追踪（绝不用光栅化），全分辨率贴图 | 路径追踪加降噪，比如带电影感打光的 Blender Cycles |
+| 编码 | H.264 或 H.265，yuv420p，CRF 18 或更低 | CRF 16，或无损帧序列 |
+
+请交母版文件。网页用的压缩版由网站自己生成，不要自己先压。
+
+## 2. 各仿真器的做法
+
+下面这些都是 2026-10-07 渲染 BEHAVIOR-1K、RoboLab、ManiGuard、RoboWits 时实际用过的，速度在一块共享的 RTX PRO 6000 上测得。
+
+### Isaac Sim / Isaac Lab（如 RoboLab）
+
+把渲染器切到路径追踪：
+
+```python
+import carb
+s = carb.settings.get_settings()
+s.set("/rtx/rendermode", "PathTracing")
+s.set("/rtx/pathtracing/spp", 16)            # 每帧每像素采样数
+s.set("/rtx/pathtracing/totalSpp", 16)
+s.set("/rtx/pathtracing/maxBounces", 6)
+s.set("/rtx/pathtracing/optixDenoiser/enabled", True)
+s.set("/rtx/pathtracing/fireflyFilter/enabled", True)
+```
+
+- 每步物理之后，先多渲几个子帧把采样攒够，再截这一帧。用 Replicator 的话就是 `rt_subframes`。
+- 只渲展示用的相机，观测相机缩小或关掉。
+- 速度：
+  - 1080p、16 spp，约 0.45 秒/帧；
+  - 1440p、32 spp，约 0.95 秒/帧。
+- 多个任务共用同一个 Isaac 进程，场景准备时间能从约 140 秒降到几秒。
+
+### OmniGibson（BEHAVIOR-1K、ManiGuard）
+
+- 路径追踪的开法同 Isaac。
+  - 想要快，可以用 RTX Real-Time 2.0（较新的 Isaac Sim 里叫 `RealTimePathTracing`），1080p 约 0.3 秒/帧。
+  - 完整路径追踪 8 spp 加 OptiX 降噪，1440p 约 1.4 秒/帧。
+- **有完整状态记录的示范**（比如 BEHAVIOR Challenge 官方的原始数据）：
+  - 用 OmniGibson 的 `DataPlaybackWrapper` 逐帧恢复状态；
+  - 两帧之间不跑物理，所以做延时摄影时只渲每隔 k 帧的那一帧，回放也完全准确。
+- **房屋尺度的场景**：
+  - 用跟拍相机，位置和朝向各做 0.5-1.5 秒的平滑，对准机器人的躯干和双手；
+  - 检查相机和机器人之间有没有墙挡着，别让机器人消失在墙后。
+- 渲染时关掉机器人自带的相机和 viewer 相机。
+
+### Genesis（如 RoboWits）
+
+- **用 Nyx**，也就是 Genesis World 1.x 的路径追踪器：
+  - 安装：`pip install genesis-world gs-nyx`；
+  - 作为相机传感器加进场景：`scene.add_sensor(NyxCameraOptions(res=(2560, 1440), ...))`；
+  - 我们用的是 FastPathTracer、128 spp、OIDN 降噪，加 Poly Haven 的 HDRI（CC0）和一盏主光；
+  - 1440p 约 0.1-0.2 秒/帧。
+- **如果套件的物理钉死在旧版 Genesis 上**：
+  1. 先用钉死的版本回放，逐帧记下每个 link 的位姿；
+  2. 再在 Genesis World 1.x 里重建同一个场景，逐帧把位姿设回去。
+  - 要核对重建后的 link 和回放是否一致。我们的误差在 1 微米以内。
+- 用 Draco 压缩的 `.glb` 网格会让 Nyx 崩溃，换成不压缩的再导出一份。
+- Genesis 的 CPU 物理在不同厂商的 CPU 上会分叉。要在录制时同类型的机器上回放，并检查是否成功。
+
+### SAPIEN / ManiSkill
+
+渲染相机用光追着色器包：
+
+```python
+env = gym.make(env_id, render_mode="rgb_array",
+               human_render_camera_configs=dict(width=2560, height=1440, shader_pack="rt"))
+```
+
+`rt-med`、`rt-fast` 用画质换速度。
+
+### MuJoCo、robosuite、PyBullet 等
+
+MuJoCo、robosuite、PyBullet 自带的渲染器都是光栅化，分辨率再高也不收。请改用 Blender Cycles 渲染：
+
+1. 回放这次运行，逐帧记录每个刚体的位姿（或直接用录下的状态）。
+2. 把同样的网格导入 Blender，逐帧设好位姿。Blender 里不跑物理。
+3. 用 GPU 上的 **Cycles** 渲染。**Eevee 和 Workbench 是光栅化，不收。**
+
+欢迎做电影感的效果。下面是 2026-10-14 重渲 RoboARC 时实际用的设置：Blender 4.5 LTS，RTX 4090 上 1440p 每帧约 1.4-6 秒。
+- **采样**：Cycles + OptiX，自适应采样（阈值 0.01，至少 64 采样），OIDN 降噪并带 albedo 和 normal 通道。
+- **布景和打光**：无缝的影棚地面，一盏大的柔光主灯，一盏较弱的顶部柔光箱和一盏轮廓光，再加一个偏暗的天空穹顶。
+- **调色**：AgX 视图变换，Medium High Contrast 风格。
+- **镜头**：缓入缓出地慢慢绕约 10 度。
+- **可选**：轻微景深和运动模糊，但机器人和目标物体必须保持清晰。
+
+## 3. 镜头和画面
+
+- **用专门的渲染相机。** 不要用 agent 的观测相机，不要用 viewer 窗口，也不要录屏。
+- **构图**：
+  - 斜上方 3/4 视角，固定机位或缓慢移动；
+  - 机器人和目标里提到的每个物体都要一直在画面里；
+  - 物体要大到看得清。
+- **关键动作放在画面中间 80% 的范围内。** 网站会把画面裁成别的比例。
+- **画面干净**：不要文字、logo、水印、UI、调试标记或坐标轴。曝光适中，窗户不能过曝发白，暗部不能死黑。
+- **同一套件风格一致**：所有题用同一种机位风格和打光。
+- **预览图**选一帧单独拿出来也看得懂的画面，一般是完成时物体都在画面里的那一帧。不要选机器人背影的特写。
+- **时长**：
+  - teaser 10-40 秒，单题视频 5-60 秒；
+  - 长任务可以加速，但速度要恒定，并写进 `meta.json`。
+
+## 4. 展示哪一次运行
+
+- **展示的必须就是评分的那道题**：同一个场景、同一个实例、同一个目标。
+- **来源**按优先级：
+  1. 参考解（oracle）；
+  2. agent 的成功 trial；
+  3. 同一任务的官方示范。
+- **不许改物理。**
+  - 按评分器（verifier）的方式重新执行录下的动作，或者回放录下的状态。只能改相机和渲染设置。
+  - 仿真器不确定的话，重新渲染结束时要再检查一次任务是否成功。
+  - **失败的回放绝不能当成功发布。**
+- **没有成功的运行？** 渲一段 6-8 秒、绕这道题自己初始场景慢慢转的镜头，再配一张预览图。不要再用通用场景图。
+
+## 5. 交付内容
+
+所有文件上传到私有的 Hugging Face 数据集 [`the review repository`](the review repository)，命令见它的 README。上传会以 pull request 的形式提交，维护者审过后合并。
+
+每道题放在 `<suite>/<task_id>/` 下：
+
+| 文件 | 内容 |
+|---|---|
+| `video.mp4` | 母版视频，规格见第 1 节 |
+| `poster.png` | 一张全分辨率的帧，按第 3 节挑选 |
+| `still.png` | 只有没有视频的题才需要：高于 1920×1080（目标 2560×1440），用光追或路径追踪渲染，要看得到机器人、每个目标物体和目标本身（例如起始和完成状态并排） |
+| `meta.json` | 见下 |
+
+整个套件另交 `<suite>/teaser.mp4` 和它的预览图。
+
+```json
+{"suite": "robolab", "task": "put_two_mugs_on_shelf", "source": "agent-success:<run>/<trial>",
+ "renderer": "isaac-rtx PathTracing spp=32 optix", "rendering": "path-tracing", "width": 2560, "height": 1440, "fps": 30,
+ "speed": 8.25, "success_check": true, "image": "rcb-robolab:0.1.1"}
+```
+
+`source` 取下面四种之一：`oracle`、`agent-success:<run>/<trial>`、`official-demo:<dataset>/<episode>`、`scene-orbit`。
+`rendering` 填 `path-tracing` 或 `ray-tracing`；`renderer` 写明渲染器和设置。
+
+## 6. 交之前先自查
+
+运行 `python3 check_media.py <suite_dir>`，需要装 ffmpeg。
+
+- **会判失败的情况**：
+  - 分辨率不高于 1920×1080、画幅不对，或帧率不达标；
+  - `rendering` 不是 `path-tracing` 或 `ray-tracing`，或者 `renderer` 里写的是光栅化渲染器（OpenGL、EGL、Eevee、Workbench 等）；
+  - 编码不是 H.264/H.265；
+  - CRF 高于 18；
+  - 缺文件或缺 `meta.json` 字段；
+  - 把失败的回放标成了成功。
+- **会给警告的情况**：尺寸低于 2560×1440 的目标、帧率不恒定、有黑边、时长超出上面的范围。
+- **自查通过后**，维护者会看过每一段 teaser，再抽查一部分单题视频，然后才上线。
+
+## 8. 补充要求一（2026-10-08）：每道题都要动起来，再交 5 个最难的
+
+1. **每道题都要有视频。** 把每道题都 replay 一遍，录下机器人做题的过程。参考解（oracle）、成功的 agent run、官方示范都可以。
+   - 只有完全没有成功运行的题，才能用静图或绕初始场景转的镜头。
+   - 要在这道题的 `meta.json` 里用 `"why_no_video"` 写明原因。
+2. **机器人要在动。** teaser 和要上首页的视频里，必须是机器人本身在动：手臂在够、手在抓、身体在走。静止的场景配上移动的相机不算。
+3. **交 5 个最难的。** 至少挑 5 道你们觉得又难又有意思的题，每道都要有机器人在动的视频。按从好到次的顺序写进 `<suite>/highlights.json`：
+
+   ```json
+   {"teaser": "<task_id>", "highlights": ["<task_id>", "<task_id>", "<task_id>", "<task_id>", "<task_id>"], "why_fewer": ""}
+   ```
+
+   - 难，指 agent 做不出来，或者需要真功夫：步骤长、要求精细、要用工具、要两只手配合、要保持平衡。
+   - teaser 必须是 highlights 里的一道。
+   - 不够 5 个的，在 `why_fewer` 里写清原因，比如"这个套件只有 4 道题"，或者"只有 3 道题有成功的运行"。
+4. **为什么要这样。** 首页滚动墙给每个套件分同样多的位置，就从 `highlights.json` 里取。套件封面用 teaser。没交 highlights 的套件，上墙的就少。
+
+`check_media.py` 现在遇到以下四种情况会判失败：
+- 没有 `highlights.json`；
+- 某个 highlight 没有机器人在动的视频；
+- highlights 不到 5 个，又没写 `why_fewer`；
+- 静图或场景环绕没写 `why_no_video`。
+
+## 9. 补充要求二（2026-10-16）：高于 1080p，只用光追或路径追踪
+
+1. **高于 1080p。** 母版必须大于 1920×1080，直接按 2560×1440（2K）或 4K 渲染。预览图和静图也一样。
+2. **只用光线追踪或路径追踪，绝不用光栅化。** 不收 OpenGL/EGL 渲染，不收 MuJoCo、robosuite、PyBullet 自带渲染器，不收 Genesis 默认相机和 SAPIEN 默认 shader，也不收 Blender 的 Eevee 和 Workbench。
+   - 仿真器没有光追的，把录下的位姿放到 Blender Cycles 里渲染。欢迎做电影感的效果，第 2 节有一套参考设置。
+3. **`meta.json` 新增 `"rendering"` 字段**，填 `"path-tracing"` 或 `"ray-tracing"`。
+4. **适用范围。** 从 2026-10-16 起的所有上传都按这个标准，重新上传也算。已经在网站上的媒体先保留，等这个套件重新上传时再替换。
+
+`check_media.py` 现在还会在以下情况判失败：
+- 视频、预览图或静图不高于 1920×1080；
+- 缺少 `rendering` 字段，或者填的值不对；
+- `renderer` 里写的是光栅化渲染器。
+
+第 6、8 节原有的检查照旧。
+
+## 7. 示例
+
+![RoboLab，1440p 路径追踪](standard_examples/robolab_example.jpg)
+
+![ManiGuard，oracle 回放，全程零违规](standard_examples/maniguard_example.jpg)

masters/ambench/cabinet_pick_place/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "cabinet_pick_place",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "CabinetPickPlace-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "cabinet_pick_place-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 2752 (120 Hz), end state c8c72a8b2031; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 1.0
+}

masters/ambench/cabinet_pick_place/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb3c06ccb4f33baf3d2df19fbe2ee40ceb3777fa9f2343f4587cac05366d930b
+size 21928607

masters/ambench/highlights.json

@@ -0,0 +1 @@
+{"teaser": "cabinet_pick_place", "highlights": ["cabinet_pick_place", "lemon_harvesting", "peg_in_hole", "rotate_valve", "pull_lever", "open_door"], "why_fewer": ""}

masters/ambench/lemon_harvesting/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "lemon_harvesting",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "LemonHarvesting-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "lemon_harvesting-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 1739 (120 Hz), end state a0426217ca46; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 1.0
+}

masters/ambench/lemon_harvesting/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b9890ccde0f2450f58ea44b48dd25a2c326755b882e82a269d24e1d6f12a563
+size 21091718

masters/ambench/ndt/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "ndt",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "NDT-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "ndt-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 1212 (120 Hz), end state bd8da2528c86; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "follows the robot from a 3/4 elevated view, smoothed over about 1 s",
+ "end_hold_s": 1.0
+}

masters/ambench/ndt/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac167f6c1d39dbbc1bd6b5eb978bc191cdca4e3cc25fb8d540103781916605e5
+size 16218179

masters/ambench/open_door/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "open_door",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "OpenDoor-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "open_door-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 395 (120 Hz), end state f778f7c7a859; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 2.0
+}

masters/ambench/open_door/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d30defaedf914639a038f07f964eb0fda6cdbf8d9a5787708b58882289d8df98
+size 4558663

masters/ambench/peg_in_hole/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "peg_in_hole",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "PegInHole-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "peg_in_hole-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 1041 (120 Hz), end state 88054491e841; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 1.0
+}

masters/ambench/peg_in_hole/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f46484371132f1d3b818b7cef303db9b63064b4708104fe0398661009669fd6
+size 24215426

masters/ambench/press_button/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "press_button",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "PressButton-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "press_button-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 1027 (120 Hz), end state fcd4de9b60a7; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 1.0
+}

masters/ambench/press_button/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4371a5c82a93a7bfc55810510833ffa451d65740d5eca2588e6f42f53651877
+size 20521162

masters/ambench/pull_lever/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "pull_lever",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "PullLever-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "pull_lever-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 899 (120 Hz), end state 1ef7b812be27; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 1.0
+}

masters/ambench/pull_lever/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af65044538726ea1a9555300bd811c867ca5e1da192e05b1f1971263c782b57c
+size 16908589

masters/ambench/push_slider/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "push_slider",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "PushSlider-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "push_slider-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 1353 (120 Hz), end state 01a85d356fea; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 1.0
+}

masters/ambench/push_slider/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:497cf40104b76743d060f38d2ead9a06a1f5ed2cc1768ddac76787635da5c444
+size 20712214

masters/ambench/rotate_valve/meta.json

@@ -0,0 +1,18 @@
+{
+ "suite": "ambench",
+ "task": "rotate_valve",
+ "source": "oracle",
+ "renderer": "isaac-rtx PathTracing (RTX Interactive, full path tracing) spp=32 (16 per update) maxBounces=6 optix denoiser, firefly filter, native 2560x1440 (Isaac Sim 5.1 / Isaac Lab, AM-Bench @ 60bf5b7)",
+ "rendering": "path-tracing",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "speed": 1.0,
+ "success_check": true,
+ "image": "eai-ambench:0.2.0",
+ "env_id": "RotateValve-Am-FAHexa-Abs-PID-Direct-v0",
+ "instance": "rotate_valve-i00 (seed 0)",
+ "replay": "the frozen oracle rows replayed as the grader does (a fresh camera-free process, the tooling of the exam's task repository PR #246); success at control step 1002 (120 Hz), end state 1b976104a883; every 4th control step rendered from the recorded states by a render-only process",
+ "camera": "3/4 elevated view along a fixed direction, aimed at the robot (smoothed over about 1 s) and the goal objects, with a slow smooth dolly that keeps both inside the central 80%",
+ "end_hold_s": 1.0
+}

masters/ambench/rotate_valve/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b647e6083a837284a0bf02cba146914d4fab0546ef8e88b61bc6f5388718b865
+size 18391319

masters/ambench/teaser.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb3c06ccb4f33baf3d2df19fbe2ee40ceb3777fa9f2343f4587cac05366d930b
+size 21928607

masters/behavior-1k/assembling_gift_baskets/meta.json

@@ -0,0 +1,32 @@
+{
+ "suite": "behavior-1k",
+ "task": "assembling_gift_baskets",
+ "source": "official-demo:behavior-1k/2026-challenge-rawdata/raw/episode_00262860.hdf5#demo_0",
+ "credit": "BEHAVIOR-1K, Stanford (2026 BEHAVIOR Challenge teleoperated demos, MIT license)",
+ "renderer": "isaac-rtx RealTimePathTracing (RTX Real-Time 2.0: ray-traced GI, reflections, AO, sampled lighting) dlss=quality render_iterations=3 hfov=68.0",
+ "width": 1920,
+ "height": 1080,
+ "fps": 30,
+ "frames": 900,
+ "speed": 34.586,
+ "seconds_per_frame": 0.172,
+ "recorded_steps": 31127,
+ "recorded_fps": 30.0,
+ "success_check": true,
+ "predicate_check_visual_only": true,
+ "success_check_how": "state playback (no re-simulation): every rendered frame restores the recorded full sim state, so the final frame is the demo's recorded final state; that state ended the recording with completion_reason=n/a, recorded terminated[-1]=True (BEHAVIOR predicate goal met at record time). predicate_check_visual_only re-evaluates PredicateGoal on the restored final state in visual-only playback (collisions off), where contact-based predicates such as ontop/inside can read false.",
+ "goal_status": {
+  "satisfied": [
+   "0",
+   "1",
+   "2",
+   "3"
+  ],
+  "unsatisfied": []
+ },
+ "playback": "DataPlaybackWrapper semantics: load_state per rendered step, transitions in order",
+ "camera": "smoothed third-person chase at 2.05 m, behind/side 3/4 view, kept inside loaded rooms and clear of furniture",
+ "image": "rcb-behavior:0.1.1 (OmniGibson 3.9.2, Isaac Sim 4.5/5.x)",
+ "created": "2026-10-07T12:19:53+0000",
+ "wall_seconds": 355.8
+}

masters/behavior-1k/assembling_gift_baskets/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7303883e4448b979c82b60ac00bf1c2af5d6d6c178df829364a83747c3c6414
+size 47695383

masters/behavior-1k/attach_a_camera_to_a_tripod/meta.json

@@ -0,0 +1,60 @@
+{
+ "suite": "behavior-1k",
+ "task": "attach_a_camera_to_a_tripod",
+ "source": "official-demo:behavior-1k/2026-challenge-rawdata/task-0035/episode_00352290.hdf5#demo_0",
+ "credit": "BEHAVIOR-1K, Stanford (2026 BEHAVIOR Challenge teleoperated demos, MIT license)",
+ "renderer": "isaac-rtx RealTimePathTracing (RTX Real-Time 2.0: ray-traced GI, reflections, AO, sampled lighting) dlss=quality render_iterations=3 hfov=68.0",
+ "width": 1920,
+ "height": 1080,
+ "fps": 30,
+ "frames": 900,
+ "speed": 4.318,
+ "seconds_per_frame": 0.094,
+ "recorded_steps": 3886,
+ "recorded_fps": 30.0,
+ "success_check": true,
+ "predicate_check_visual_only": true,
+ "success_check_how": "state playback (no re-simulation): every rendered frame restores the recorded full sim state, so the final frame is the demo's recorded final state; that state ended the recording with completion_reason=n/a, recorded terminated[-1]=True (BEHAVIOR predicate goal met at record time). predicate_check_visual_only re-evaluates PredicateGoal on the restored final state in visual-only playback (collisions off), where contact-based predicates such as ontop/inside can read false.",
+ "goal_status": {
+  "satisfied": [
+   "0"
+  ],
+  "unsatisfied": []
+ },
+ "state_layout_skipped_objects": null,
+ "poster_camera": {
+  "method": "box-occlusion visibility",
+  "score": 3.555,
+  "goal_objects_visible": "2/2",
+  "robot_height_fraction": 0.745,
+  "wall_fraction": 0.0,
+  "blown": 0.0008,
+  "eye": [
+   -6.01,
+   2.36,
+   2.05
+  ],
+  "target": [
+   -3.54,
+   1.98,
+   0.85
+  ],
+  "candidates": 1,
+  "tag": "chase",
+  "search_seconds": 2.8
+ },
+ "playback": "DataPlaybackWrapper semantics: load_state per rendered step, transitions in order",
+ "camera": "smoothed third-person chase at 2.05 m, behind/side 3/4 view, kept inside loaded rooms and clear of furniture",
+ "scene": "challenge full-scene template + task rooms",
+ "image": "stanfordvl/behavior:3.9.2@sha256:2702900192741904e69a573aa7869ea51c592586f9791050796efc37faa893fb (official BEHAVIOR image, loaded from the offloaded tar); OmniGibson 3.9.2, Isaac Sim 5.1.0; BEHAVIOR assets 3.9.0, robot assets 3.8.2",
+ "created": "2026-10-08T08:45:21+0000",
+ "wall_seconds": 165.1,
+ "host": "laptop (mll-laptop-1, RTX 4090 Laptop 16 GB, container rcbmedia-behavior-laptop)",
+ "recorded_with": {
+  "omnigibson": "3.7.0-alpha",
+  "bddl": "3.7.0a0",
+  "og_dataset": "1.2.0rc21"
+ },
+ "episode_selection": "near-median size (10.6 MB; task median 10.6 MB, 200 public episodes), 5 rollbacks, recorded success flag true; 1 candidate(s) probed",
+ "poster": "final recorded state, same renderer, camera chosen by goal-object visibility, 8 render iterations"
+}

masters/behavior-1k/attach_a_camera_to_a_tripod/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:603ca763ea8c86cbdc6ea2a38b65b16a0695954397e748e1e9bdaf6a69fb5a8c
+size 26044013

masters/behavior-1k/boxing_books_up_for_storage/meta.json

@@ -0,0 +1,29 @@
+{
+ "suite": "behavior-1k",
+ "task": "boxing_books_up_for_storage",
+ "source": "official-demo:behavior-1k/2026-challenge-rawdata/raw/episode_00230610.hdf5#demo_0",
+ "credit": "BEHAVIOR-1K, Stanford (2026 BEHAVIOR Challenge teleoperated demos, MIT license)",
+ "renderer": "isaac-rtx RealTimePathTracing (RTX Real-Time 2.0: ray-traced GI, reflections, AO, sampled lighting) dlss=quality render_iterations=3 hfov=68.0",
+ "width": 1920,
+ "height": 1080,
+ "fps": 30,
+ "frames": 900,
+ "speed": 29.857,
+ "seconds_per_frame": 0.101,
+ "recorded_steps": 26871,
+ "recorded_fps": 30.0,
+ "success_check": true,
+ "predicate_check_visual_only": true,
+ "success_check_how": "state playback (no re-simulation): every rendered frame restores the recorded full sim state, so the final frame is the demo's recorded final state; that state ended the recording with completion_reason=n/a, recorded terminated[-1]=True (BEHAVIOR predicate goal met at record time). predicate_check_visual_only re-evaluates PredicateGoal on the restored final state in visual-only playback (collisions off), where contact-based predicates such as ontop/inside can read false.",
+ "goal_status": {
+  "satisfied": [
+   "0"
+  ],
+  "unsatisfied": []
+ },
+ "playback": "DataPlaybackWrapper semantics: load_state per rendered step, transitions in order",
+ "camera": "smoothed third-person chase at 2.05 m, behind/side 3/4 view, kept inside loaded rooms and clear of furniture",
+ "image": "rcb-behavior:0.1.1 (OmniGibson 3.9.2, Isaac Sim 4.5/5.x)",
+ "created": "2026-10-07T12:23:13+0000",
+ "wall_seconds": 194.3
+}

masters/behavior-1k/boxing_books_up_for_storage/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3361df463327a020351eb9689452cd74667ba1e40a054fd1a07b4c84f6824cba
+size 33197174

masters/behavior-1k/boxing_food_after_dinner/meta.json

@@ -0,0 +1,66 @@
+{
+ "suite": "behavior-1k",
+ "task": "boxing_food_after_dinner",
+ "source": "official-demo:behavior-1k/2026-challenge-rawdata/task-0067/episode_00670010.hdf5#demo_1",
+ "credit": "BEHAVIOR-1K, Stanford (2026 BEHAVIOR Challenge teleoperated demos, MIT license)",
+ "renderer": "isaac-rtx RealTimePathTracing (RTX Real-Time 2.0: ray-traced GI, reflections, AO, sampled lighting) dlss=quality render_iterations=3 hfov=68.0",
+ "width": 1920,
+ "height": 1080,
+ "fps": 30,
+ "frames": 900,
+ "speed": 7.996,
+ "seconds_per_frame": 0.21,
+ "recorded_steps": 7196,
+ "recorded_fps": 30.0,
+ "success_check": true,
+ "predicate_check_visual_only": false,
+ "success_check_how": "state playback (no re-simulation): every rendered frame restores the recorded full sim state, so the final frame is the demo's recorded final state; that state ended the recording with completion_reason=task_success, recorded terminated[-1]=True (BEHAVIOR predicate goal met at record time). predicate_check_visual_only re-evaluates PredicateGoal on the restored final state in visual-only playback (collisions off), where contact-based predicates such as ontop/inside can read false.",
+ "goal_status": {
+  "satisfied": [
+   "0",
+   "3",
+   "4"
+  ],
+  "unsatisfied": [
+   "1",
+   "2"
+  ]
+ },
+ "state_layout_skipped_objects": null,
+ "poster_camera": {
+  "method": "box-occlusion visibility",
+  "score": 2.997,
+  "goal_objects_visible": "7/7",
+  "robot_height_fraction": 0.627,
+  "wall_fraction": 0.04,
+  "blown": 0.0112,
+  "eye": [
+   5.15,
+   -0.02,
+   2.11
+  ],
+  "target": [
+   6.98,
+   -1.86,
+   0.93
+  ],
+  "candidates": 151,
+  "tag": "t0f0r2.6a135h2.1",
+  "search_seconds": 9.4
+ },
+ "playback": "DataPlaybackWrapper semantics: load_state per rendered step, transitions in order",
+ "camera": "smoothed third-person chase at 2.05 m, behind/side 3/4 view, kept inside loaded rooms and clear of furniture",
+ "scene": "challenge full-scene template + task rooms",
+ "image": "stanfordvl/behavior:3.9.2@sha256:2702900192741904e69a573aa7869ea51c592586f9791050796efc37faa893fb (official BEHAVIOR image, loaded from the offloaded tar); OmniGibson 3.9.2, Isaac Sim 5.1.0; BEHAVIOR assets 3.9.0, robot assets 3.8.2",
+ "created": "2026-10-08T11:15:59+0000",
+ "wall_seconds": 447.3,
+ "host": "laptop (mll-laptop-1, RTX 4090 Laptop 16 GB, container rcbmedia-behavior-laptop)",
+ "recorded_with": {
+  "omnigibson": "3.8.0",
+  "bddl": "3.7.0",
+  "behavior-1k-assets": "3.7.2rc1",
+  "omnigibson-robot-assets": "3.8.2"
+ },
+ "episode_selection": "near-median size (48.8 MB; task median 48.8 MB, 200 public episodes), 10 rollbacks, recorded success flag true; 1 candidate(s) probed",
+ "poster": "final recorded state, same renderer, camera chosen by goal-object visibility, 8 render iterations"
+}

masters/behavior-1k/boxing_food_after_dinner/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4594ddb8965f4de52f43307fdafc86c894716e04f707fe55a76f00967280dc1
+size 37854882

masters/behavior-1k/bringing_in_wood/meta.json

@@ -0,0 +1,56 @@
+{
+ "suite": "behavior-1k",
+ "task": "bringing_in_wood",
+ "source": "official-demo:behavior-1k/2026-challenge-rawdata/task-0015/episode_00150330.hdf5#demo_0",
+ "credit": "BEHAVIOR-1K, Stanford (2026 BEHAVIOR Challenge teleoperated demos, MIT license)",
+ "renderer": "isaac-rtx RealTimePathTracing (RTX Real-Time 2.0: ray-traced GI, reflections, AO, sampled lighting) dlss=quality render_iterations=3 hfov=68.0",
+ "width": 1920,
+ "height": 1080,
+ "fps": 30,
+ "frames": 900,
+ "speed": 17.429,
+ "seconds_per_frame": 0.157,
+ "recorded_steps": 15686,
+ "recorded_fps": 30.0,
+ "success_check": true,
+ "predicate_check_visual_only": false,
+ "success_check_how": "state playback (no re-simulation): every rendered frame restores the recorded full sim state, so the final frame is the demo's recorded final state; that state ended the recording with completion_reason=n/a, recorded terminated[-1]=True (BEHAVIOR predicate goal met at record time). predicate_check_visual_only re-evaluates PredicateGoal on the restored final state in visual-only playback (collisions off), where contact-based predicates such as ontop/inside can read false.",
+ "goal_status": {
+  "satisfied": [],
+  "unsatisfied": [
+   "0"
+  ]
+ },
+ "state_layout_skipped_objects": [
+  "charcoal_grill_crdaeu_0"
+ ],
+ "poster_camera": {
+  "method": "box-occlusion visibility",
+  "score": 3.05,
+  "goal_objects_visible": "3/3",
+  "robot_height_fraction": 0.45,
+  "wall_fraction": 0.33,
+  "blown": 0.0,
+  "eye": [
+   2.64,
+   4.71,
+   1.7
+  ],
+  "target": [
+   -0.9,
+   1.18,
+   0.9
+  ],
+  "candidates": 130,
+  "tag": "f2r5.0a45h1.7",
+  "search_seconds": 2.4
+ },
+ "playback": "DataPlaybackWrapper semantics: load_state per rendered step, transitions in order",
+ "camera": "smoothed third-person chase at 2.05 m, behind/side 3/4 view, kept inside loaded rooms and clear of furniture",
+ "scene": "challenge full-scene template + task rooms",
+ "image": "rcb-behavior:0.1.1 (a workstation image id sha256:1f09f0e6f4fba5eac0fe9e57e24a3c04255526593e9b99c77e8e123284dc7976; lab copy sha256:ccc8fc28138df9416a3eb45bc138c1aab3a13fb610abe715d67e0130971743e6, built from stanfordvl/behavior:3.9.2 + images/behavior/Dockerfile); OmniGibson 3.9.2, Isaac Sim 5.1.0; BEHAVIOR assets 3.9.0, robot assets 3.8.2",
+ "created": "2026-10-07T17:36:48+0000",
+ "wall_seconds": 254.0,
+ "host": "a workstation (a workstation, RTX 4090 24 GB, shared host)",
+ "poster": "final recorded state, same renderer, camera chosen by goal-object visibility, 8 render iterations"
+}

masters/behavior-1k/bringing_in_wood/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac02861463a36e4c91408ef0cfa5837db36ceb5499956890fbf3f03d22a01828
+size 65717061