Update completed Mingru MuJoCo runs with full campaign state and provenance (part 20)

Full run folders, inputs, prompts, generation logs, outputs, contracts, evaluator protocols, historical attempts and SHA256 manifest. Snapshot: 2026-11-12T08:06:46.141011+00:00

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/evaluator_protocol.json

@@ -0,0 +1,50 @@
+{
+  "track": "demos",
+  "engine": "mujoco",
+  "engine_version": "4.0.0",
+  "protocols": {
+    "all": {
+      "engine_version": "4.0.0",
+      "episodes": {
+        "file": "configs/demos.json",
+        "fields": [
+          "id",
+          "moved",
+          "relation",
+          "target"
+        ],
+        "count": 88,
+        "sha256": "f319e80520cccb0cb3da8c329f879d6c1367f7c91839661ea0a816853e26540c"
+      },
+      "held_out": "numbers only: the frozen rule was not met on the controls",
+      "depth": "derived from stereo, numbers only",
+      "held_out_view": {
+        "rule": "the camera the agent is not given",
+        "camera": "wrist",
+        "given": [
+          "ext1",
+          "ext2"
+        ]
+      },
+      "given_views": {
+        "rule": "the cameras the agent was given, at the held-out view's film steps: PSNR, SSIM and derived stereo depth by the held-out formula (robot masked, calibration offset applied), over the whole frame (scene), over a cube around the moved object's logged position at the steps it is free of the gripper, and over a box under the release point (object)",
+        "moved_half_m": "half the held opening (at least the margin) plus the margin",
+        "margin_m": 0.01,
+        "target_half_m": 0.05,
+        "target_depth_m": 0.1,
+        "scored": false,
+        "engines": "both; Isaac Sim films the given views in its render container, as the held-out view"
+      },
+      "free_space_before_the_grasp": "the fingers and pads may touch the moved object before they close; any other link may not",
+      "pass": "contact points, free space and replay outcome all pass",
+      "limits": {
+        "position_m": 0.03,
+        "width_m": 0.008,
+        "free_space_force_n": 5.0,
+        "contact_listed_n": 1.0,
+        "minimum_rise_fraction": 0.5,
+        "resting_m": 0.02
+      }
+    }
+  }
+}

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/input/cameras.json

@@ -0,0 +1,70 @@
+{
+ "ext1": {
+  "fx": 522.4622802734375,
+  "fy": 522.4622802734375,
+  "cx": 640.4918212890625,
+  "cy": 353.2307434082031,
+  "width": 1280,
+  "height": 720,
+  "pose_camera_to_base": [
+   [
+    0.8919446480746538,
+    -0.40554250746042825,
+    0.19992503448465607,
+    0.24817175873665903
+   ],
+   [
+    -0.41578764657202255,
+    -0.5619640067307263,
+    0.7150643943710575,
+    -0.4977961989764716
+   ],
+   [
+    -0.17763833406413518,
+    -0.7209242191672138,
+    -0.6698603529759484,
+    0.5158103419831875
+   ],
+   [
+    0.0,
+    0.0,
+    0.0,
+    1.0
+   ]
+  ]
+ },
+ "ext2": {
+  "fx": 522.9500732421875,
+  "fy": 522.9500732421875,
+  "cx": 639.2383422851562,
+  "cy": 352.5005798339844,
+  "width": 1280,
+  "height": 720,
+  "pose_camera_to_base": [
+   [
+    -0.965118677369552,
+    -0.17445502920227396,
+    0.19522136506663584,
+    0.5056940164360756
+   ],
+   [
+    -0.26133437816137495,
+    0.6869739800123663,
+    -0.6780642252596554,
+    0.41623301203765845
+   ],
+   [
+    -0.015820283924583525,
+    -0.7054305022977125,
+    -0.7086025155505328,
+    0.7893579556764505
+   ],
+   [
+    0.0,
+    0.0,
+    0.0,
+    1.0
+   ]
+  ]
+ }
+}

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/input/ext1.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30c538a7537b05c4400a1222f0704049c4b8b68b629e461661bd541893dd4600
+size 2497282

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/input/ext2.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11918f23092fd8db286a6340909a2cf27ae6d51304f3c1e5a8b8324a3934c704
+size 2496567

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/input/instruction.txt

@@ -0,0 +1 @@
+Put the marker in the bowl

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/input/timestamps.json

@@ -0,0 +1 @@
+{"ext1": [0.002, 0.068, 0.135, 0.202, 0.285, 0.353, 0.419, 0.485, 0.569, 0.635, 0.702, 0.769, 0.852, 0.919, 0.985, 1.052, 1.135, 1.202, 1.269, 1.352, 1.419, 1.485, 1.552, 1.619, 1.685, 1.769, 1.835, 1.902, 1.985, 2.052, 2.119, 2.185, 2.252, 2.335, 2.402, 2.469, 2.535, 2.619, 2.685, 2.752, 2.819, 2.885, 2.952, 3.035, 3.102, 3.169, 3.235, 3.302, 3.386, 3.452, 3.519, 3.602, 3.669, 3.736, 3.802, 3.869, 3.952, 4.019, 4.086, 4.169, 4.236, 4.302, 4.369, 4.435, 4.519, 4.586, 4.652, 4.719, 4.802, 4.869, 4.936, 5.019, 5.086, 5.152, 5.219, 5.302, 5.369, 5.436, 5.502, 5.586, 5.636, 5.719, 5.786, 5.852, 5.936, 6.002, 6.069, 6.136, 6.219, 6.286, 6.352, 6.419, 6.502, 6.569, 6.636, 6.703, 6.786, 6.853, 6.919, 6.986, 7.069, 7.136, 7.202, 7.269, 7.336, 7.419, 7.486, 7.553, 7.636, 7.703, 7.769, 7.836, 7.903, 7.986], "ext2": [0.0, 0.067, 0.133, 0.216, 0.283, 0.35, 0.416, 0.483, 0.567, 0.633, 0.7, 0.783, 0.85, 0.917, 0.983, 1.067, 1.133, 1.2, 1.267, 1.333, 1.4, 1.483, 1.55, 1.633, 1.7, 1.767, 1.833, 1.9, 1.967, 2.05, 2.117, 2.183, 2.267, 2.333, 2.4, 2.467, 2.534, 2.6, 2.684, 2.75, 2.817, 2.884, 2.95, 3.033, 3.1, 3.167, 3.25, 3.317, 3.384, 3.45, 3.517, 3.6, 3.667, 3.734, 3.8, 3.884, 3.95, 4.017, 4.084, 4.15, 4.234, 4.3, 4.367, 4.45, 4.517, 4.584, 4.65, 4.734, 4.8, 4.867, 4.934, 5.017, 5.084, 5.15, 5.217, 5.284, 5.367, 5.434, 5.5, 5.567, 5.651, 5.717, 5.784, 5.85, 5.934, 6.0, 6.084, 6.151, 6.217, 6.284, 6.351, 6.434, 6.5, 6.567, 6.65, 6.717, 6.784, 6.85, 6.917, 7.001, 7.067, 7.134, 7.201, 7.267, 7.351, 7.417, 7.484, 7.55, 7.617, 7.7, 7.767, 7.851, 7.917, 7.984]}

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/run.json

@@ -0,0 +1,124 @@
+{
+  "track": "demos",
+  "engine": "mujoco",
+  "item": "IRIS+ef107c48+2023-03-09-19h-51m-39s",
+  "run_id": "20261112T072142Z",
+  "campaign": "mingru-mujoco",
+  "cli": "opencode",
+  "model": "opencode-go/glm-5.3-flash",
+  "model_reasoning_effort": "default",
+  "budget_seconds": 2700,
+  "told_budget_seconds": 2700,
+  "stage": 1,
+  "continued_from": null,
+  "contract": {
+    "name": "demos-mujoco",
+    "sha256": "62bafbc75aef84ee7cc9ff19cfa81dae26a0da023f63a84e9e84517afa3c8770",
+    "track": "demos",
+    "engine": "mujoco",
+    "engine_version": "4.0.0",
+    "image": "simready-generation:latest",
+    "blender_version": "5.2.3",
+    "inputs": [
+      "instruction.txt",
+      "ext1.mp4",
+      "ext2.mp4",
+      "timestamps.json",
+      "cameras.json (both exterior poses and intrinsics)"
+    ],
+    "withheld": [
+      "groundtruth/: the wrist video and its calibration, the stereo videos, the robot log (trajectory.h5), the end-state label, the measured ground truth",
+      "camera serials and provenance/",
+      "the frozen protocol of the track and engine (data/benchmark/frozen/<track>/<engine>.json)",
+      "other runs"
+    ],
+    "outputs": [
+      "scene.xml (bodies moved, target, table; only target when the table is the target)",
+      "assets/",
+      "objects.json",
+      "README.md"
+    ],
+    "self_check": "python /work/tools/check_package.py",
+    "budget": {
+      "told": "the first of the session's cumulative budgets, in wall-clock seconds from container start",
+      "resume": "each later budget resumes the same conversation for the rest of its time; each budget is its own run",
+      "enforcement": "the runner kills the container at the deadline; the output as it stands is the result"
+    },
+    "refinement": null,
+    "isolation": {
+      "container": "Docker: read-only input (and tools), writable output and scratch, temporary credentials; no repository, benchmark ground truth or evaluator files",
+      "network": "an internal network whose only exit is an HTTPS CONNECT proxy (runtime/egress_proxy.py)",
+      "egress_hosts": [
+        "api.anthropic.com",
+        "platform.claude.com",
+        "chatgpt.com",
+        "auth.openai.com",
+        "opencode.ai",
+        "mcp.exa.ai",
+        "search.parallel.ai"
+      ]
+    },
+    "protocol": {
+      "file": "data/benchmark/frozen/demos/mujoco.json",
+      "sha256": "c08e39588e6ae87b849ae2efda11102b69f3106e17d580c8f8f5c7ca0e80c05f"
+    }
+  },
+  "engine_version": "4.0.0",
+  "image": "simready-generation:latest",
+  "image_id": "sha256:5ac8fc30775a68f33a3df6b49a859161320a53170813ffb0c495e2e6fc2fb81d",
+  "required_refinement_rounds": null,
+  "gpu_uuid": null,
+  "input_snapshot": "input",
+  "tools": "tools",
+  "evaluator_protocol": "evaluator_protocol.json",
+  "output": "output",
+  "log": "generation.jsonl",
+  "status": "budget_expired",
+  "agent_cli": {
+    "cli": "opencode",
+    "name": "OpenCode",
+    "version": "1.19.5",
+    "binary": "/opt/opencode/bin/opencode in both generation images",
+    "binary_sha256": "7df4e4136301f16589b0f30bf91a61380fca5f8ef22dd36570f5ebe1d4ec802c",
+    "login": "host OpenCode V2 auth export/import API, or legacy ~/.local/share/opencode/auth.json",
+    "billing": "subscription (OpenCode Go)",
+    "tools": "every tool but webfetch (denied by its permission config, so not offered); websearch on (Exa or Parallel)",
+    "hosts": [
+      "opencode.ai",
+      "mcp.exa.ai",
+      "search.parallel.ai"
+    ],
+    "model": "opencode-go/glm-5.3-flash",
+    "image_input": null
+  },
+  "started_at": "2026-11-12T07:21:43.035792+00:00",
+  "model_answered": true,
+  "exit_code": 137,
+  "credential_refresh_persistence": "unchanged",
+  "finished_at": "2026-11-12T08:06:43.332663+00:00",
+  "cli_session_id": "ses_e316ca353ffe4a6ikr51dQ4jJb",
+  "api_retries": 0,
+  "rate_limit_events": 0,
+  "stage_seconds": 2700.264,
+  "elapsed_seconds": 2700.264,
+  "cost": {
+    "status": "measured",
+    "source": "OpenCode session database (cumulative)",
+    "tokens": {
+      "input": 183549,
+      "cache_read": 4743680,
+      "cache_write": 0,
+      "output": 80838,
+      "reasoning": 62653,
+      "total": 5008067
+    },
+    "turns": 89,
+    "usd": 0.21026174999999991,
+    "billing": "subscription (OpenCode Go)",
+    "usd_basis": "reported by OpenCode at its per-token rates; billed as subscription (OpenCode Go)",
+    "wall_seconds": 2700.264,
+    "budget_seconds": 2700,
+    "stopped_at_budget": true
+  },
+  "scene_present": false
+}

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/submitted_prompt.txt

@@ -0,0 +1,53 @@
+You are rebuilding the scene of a recorded robot demonstration as a simulation-ready MuJoCo scene. A real
+Franka Panda with a Robotiq 2F-85 gripper did this task:
+
+    Put the marker in the bowl
+
+Inputs (read-only, /work/input):
+- instruction.txt: the task above.
+- ext1.mp4 and ext2.mp4: the two fixed exterior cameras, 1280x720, filming at the same time from different places.
+- timestamps.json: seconds per video frame for each video, on one clock from the first frame. The videos' containers
+  say 60 fps, which is wrong: frames are about 67 ms apart, so use these times.
+- cameras.json: "ext1" and "ext2" each have "pose_camera_to_base", a 4x4 matrix, and intrinsics. A camera's +Z looks forward, +X is right and +Y is down; u = fx*x/z+cx, v = fy*y/z+cy.
+The calibration is metric: a point seen by both exterior cameras can be triangulated to its real position, which gives
+real sizes, heights and distances.
+Look at frames of both videos before modelling: extract them with imageio_ffmpeg in Python, and open them with the read tool. Listing files
+does not count as viewing.
+
+The world frame is the robot's base: metres, kilograms, +Z up, +X forward from the robot, origin at the centre of the arm's base where it
+is mounted. Use the cameras' poses to place things. You are not given the robot's joint angles or gripper pose.
+
+Rebuild what the robot's task depends on, where it really is:
+- "moved": the object the robot picks up. One rigid body, free to move, resting where it really starts, with mass, friction
+  and collision geometry of its real shape and size. The fingers must be able to close on it as they really did.
+- "target": what the object is put in or on. One rigid body or static, with collision geometry of its real shape: if it is
+  a vessel, its cavity must be open, since a convex shape would be solid.
+- "table": the surface they stand on, static, at its real height and extent. If the object is put on the table itself,
+  the table is the target: build it once, as "target", and leave out "table".
+- anything else the arm could reach in the real scene, static and named, only if it is really there.
+Give things their real appearance as well. Do not model the robot or anything that moves without contact.
+The arm stands at the origin: keep the space the real arm moved through, above the table, free of colliders.
+
+Output (/work/output):
+- scene.xml: a self-contained MJCF file for MuJoCo 4.0, loadable by mujoco.MjModel.from_xml_path. Top-level bodies named "moved"
+  (with one <freejoint/>), "target" and "table" (only "target" when the table is the target), plus any others, every one named. "moved" and "target" have no child bodies.
+  A body with no joint is static. No <actuator>, <equality>, <keyframe>, <plugin>, mocap, gravcomp, robot or camera tricks.
+  Visible and collision geometry may be separate: a geom with contype="0" conaffinity="0" is visual only, and give it mass="0" if
+  a collider carries the mass. MuJoCo collides a mesh as its convex hull, so build a concave object from several geoms.
+- assets/: every mesh and texture scene.xml references, by relative path. Generate meshes yourself (OBJ or STL), in Python, Blender or any other tool.
+- objects.json: {"objects": [{"id": "moved", "label": "<its name>"}, {"id": "target", "label": "<its name>"}, {"id": "table", "label": "table"}]}; when the table is the target, list it once,
+  as {"id": "target", "label": "table"}.
+- README.md: which frames you used, the estimates you made and their uncertainty.
+
+Runtime: python runs Python with MuJoCo, NumPy, SciPy, Pillow, imageio-ffmpeg, trimesh, Open3D, OpenCV, scikit-image,
+PyMeshLab, manifold3d, usd-core, shapely, pandas, matplotlib and scikit-learn (pip list shows all); ffmpeg is on PATH.
+MUJOCO_GL=osmesa supports headless rendering with
+mujoco.Renderer. Test that the scene loads and rests, render it from ext1's and ext2's calibration, and compare with real frames.
+
+Reconstruct what the videos show; do not replace objects with convenient templates if their shape matters. Mass and friction are
+estimates. Do not use private data. Download nothing from the internet (no files, models, assets, datasets or packages),
+and do not look up this demonstration, its videos or the dataset it comes from. Do not read authentication files. You may
+use any tools, including web search, and the software installed here, including Blender (`blender`).
+
+Before finishing run `python /work/tools/check_package.py`: it checks the format only, not quality, and prints what to fix.
+Budget: 2700 seconds of wall-clock time; save a valid scene well before the end. Write code and notes in English.

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/token_usage.json

@@ -0,0 +1,10 @@
+{
+ "sessions": 1,
+ "input": 183549,
+ "output": 18185,
+ "reasoning": 62653,
+ "cache_read": 4743680,
+ "cache_write": 0,
+ "cost": 0.21026174999999991,
+ "assistant_messages": 89
+}

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/tools/check_package.py

@@ -0,0 +1,49 @@
+"""Check a demonstration package's format, not its quality. Run by the agent: python /work/tools/check_package.py [folder]"""
+import json
+from pathlib import Path
+import sys
+
+import mujoco
+
+sys.path.insert(0, str(Path(__file__).parent))
+from mjcf_facts import facts  # noqa: E402
+from rules import violations  # noqa: E402
+
+folder = Path(sys.argv[1] if len(sys.argv) > 1 else "/work/output")
+issues = []
+manifest = folder/"objects.json"
+objects = json.loads(manifest.read_text())["objects"] if manifest.is_file() else []
+# When the object is put on the table itself, the table is built once, as "target".
+table_is_target = any(o.get("id") == "target" and str(o.get("label", "")).strip().lower() == "table" for o in objects)
+required = ("moved", "target") if table_is_target else ("moved", "target", "table")
+scene = folder/"scene.xml"
+if not scene.is_file():
+    issues.append("no scene.xml")
+else:
+    issues += violations(facts(scene))
+    try:
+        model = mujoco.MjModel.from_xml_path(str(scene))
+    except ValueError as error:
+        model = None
+        issues.append(f"does not load: {error}")
+    if model:
+        top = [b for b in range(1, model.nbody) if model.body_parentid[b] == 0]
+        names = [model.body(b).name for b in top]
+        issues += ["a top-level body has no name"] if "" in names else []
+        issues += [f'no top-level body "{n}"' for n in required if n not in names]
+        for name in ("moved", "target"):
+            if name in names:
+                body = model.body(name).id
+                issues += [f'"{name}" has a child body'] if any(model.body_parentid[b] == body for b in range(model.nbody)) else []
+                issues += [f'"{name}" has no geoms'] if not model.body_geomnum[body] else []
+        if "moved" in names:
+            body = model.body("moved").id
+            if model.body_jntnum[body] != 1 or model.jnt_type[model.body_jntadr[body]] != mujoco.mjtJoint.mjJNT_FREE:
+                issues.append('"moved" needs one free joint')
+if not manifest.is_file():
+    issues.append("no objects.json")
+else:
+    listed = [o["id"] for o in objects]
+    issues += [f'objects.json lists no "{n}"' for n in ("moved", "target") if n not in listed]
+print("\n".join(issues) or "format ok")
+sys.exit(1 if issues else 0)

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/tools/mjcf_facts.py

@@ -0,0 +1,21 @@
+"""What a submitted MJCF says, read from its text before anything compiles. Standard library only: the agent's checker uses it too."""
+import xml.etree.ElementTree as ET
+
+from rules import ROBOT_NAME
+
+
+def trees(path):
+    """scene.xml and everything it includes."""
+    root = ET.parse(path).getroot()
+    return [root, *(tree for node in root.iter("include") for tree in trees(path.parent/node.get("file")))]
+
+
+def facts(path):
+    roots = trees(path)
+    nodes = lambda tag: [node for root in roots for node in root.iter(tag)]
+    bodies = nodes("body")
+    constraints = [child for node in nodes("equality") for child in node]
+    return {"actuators": [child.tag for node in nodes("actuator") for child in node], "keyframes": [node.tag for node in nodes("key")],
+            "welds": [f'{n.get("body1", n.get("name", n.tag))} to {n.get("body2", "world")}' for n in constraints], "plugins": [node.tag for node in nodes("plugin")],
+            "mocap": [b.get("name") for b in bodies if b.get("mocap") == "true"], "gravity_off": [b.get("name") for b in bodies if float(b.get("gravcomp", 0)) != 0],
+            "robot_names": [n for n in (x.get("name") for x in [*bodies, *nodes("joint")]) if n and ROBOT_NAME.search(n)]}

data/runs/demos/mujoco/IRIS+ef107c48+2023-03-09-19h-51m-39s/20261112T072142Z/tools/rules.py

@@ -0,0 +1,16 @@
+"""What a demonstration scene may not contain: the robot is not part of the scene, and only contact moves things.
+
+Pure rules over facts read from the file, so MuJoCo (host) and Isaac (container, without this package) share them.
+"""
+import re
+
+ROBOT_NAME = re.compile(r"(^|[/_])(panda|franka|robotiq|2f.?85|fr3|gripper|end.?effector|finger|robot|manipulator)", re.I)
+FACTS = ("actuators", "keyframes", "welds", "mocap", "robot_names", "animated", "scripts", "initial_velocity", "gravity_off", "plugins")
+RULES = {"actuators": "actuators or drives", "keyframes": "keyframes", "welds": "welds or fixed joints", "mocap": "mocap or kinematic bodies",
+         "robot_names": "a submitted robot", "animated": "scripted motion: time-sampled transforms", "scripts": "scripted motion: graphs or scripts",
+         "initial_velocity": "a body started in motion", "gravity_off": "a body exempt from gravity", "plugins": "plugins"}
+
+
+def violations(facts):
+    """Every broken rule, each with what broke it. Reject on any: a scene that passes the check can only move by contact."""
+    return [f"{RULES[key]}: {', '.join(map(str, facts[key][:5]))}" for key in FACTS if facts.get(key)]