--- license: apache-2.0 task_categories: - robotics tags: - LeRobot - SIMPLE - humanoid - unitree-g1 - loco-manipulation - box-carry - multi-robot - DART configs: - config_name: default data_files: data/*/*.parquet --- # SimFactoryBoxCarryDualDART-v1 Two-humanoid box-carrying demonstrations for the Unitree G1 with Dex3 hands, recorded with DART-style perturbations: while the scripted expert carries the box, noise is injected into what the robots *execute*, so the episodes visit off-nominal states (robots drifting apart along the box, the box sagging in the grip, low lifts) and show the expert's corrections. The `action` labels stay the expert's clean targets. Same task, layout and cameras as [`SimFactoryBoxCarryDual-v1-mujoco`](https://huggingface.co/datasets/randing2000/SimFactoryBoxCarryDual-v1-mujoco); rendered in MuJoCo. **Task.** Two G1s face each other at the two ends of a long box lying across a table. Each grips its end with both hands; together they lift it, side-step to a second table, set it down and let go. Each recorded run gives two LeRobot episodes, one per robot, in the single-robot format: that robot's own head camera and its own `states` / `action`, with a role instruction as the task: - robot A: "You hold the left end of the long box. Lift it together with your partner, side-step to the other table and put it down." - robot B: "You hold the right end of the long box. Lift it together with your partner, side-step to the other table and put it down." Episodes 2k (robot A) and 2k+1 (robot B) come from the same run; `environment_config.role` and `environment_config.pair` record the link. Both episodes of a run carry the same side-camera video. A preview of one run is in [`previews/d-7001.mp4`](previews/d-7001.mp4): robot A's head camera, robot B's head camera and the side camera, with a strip showing the phase and the live perturbation (red while DART is active). ## Contents | Item | Value | |---|---| | Episodes | 1,236 (618 runs x 2 robots; 35 to 55 s each, 14.6 h in total) | | Frames | 2,635,328 at 50 fps | | Cameras | `observation.images.egocentric`: the robot's own G1 `head_stereo_left`; `observation.images.side`: a fixed scene camera beside the carry path. Both 640x360, h264 | | `states` / `action` | 32-d / 36-d, same semantics as SIMPLE (see `meta/modality.json`); `action` = the clean expert targets | | `action_executed` | 36-d, same layout as `action`: what the perturbed robot actually executed (label arm targets + the DART arm offset, the noisy AMO height / vx / vy / target yaw). Equal to `action` where no noise is applied | | `observation.dart_active` | 1 while noise is applied at that step (62 % of the rows), else 0 | | Format | LeRobot codebase v2.1, `video_path` = `videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4`; 2 chunks of up to 1,000 episodes | ## Perturbations DART noise level sigma = 0.3 (`environment_config.dart` holds sigma, seed, scales, phases). Ornstein-Uhlenbeck noise (correlation time 0.5 s) plus short bursts (0.3 to 0.8 s), applied only to the executed targets, by phase: - **Lift** (`hug_lift`): one hand offset shared by all four hands (each arm's joint offsets from its wrist Jacobian, so the held box moves rigidly), bursts pointing down (lift under-shoot), plus AMO height noise. - **Side-step** (`carry_*`): `spacing` (the same backward/forward velocity offset on both robots, so they drift apart or together along the box), shared and per-robot vx / vy, target-yaw and height noise, and box force pulses: horizontal (6 N, 0.2 s) and downward (6 N, 0.3 to 0.6 s, the box sags in the grip). - **Set-down**: AMO height noise. - The 1 s verified hold after the lift is left clean; the grasp and the set-down crouch get no hand noise (calibration: the full-hand dual grip loses the box under any relative hand jitter there). The expert reacts through its normal feedback: each robot steps back to its stance line across the transport, the team re-lifts a sagging box, the lift is completed before the side-step. ## Generation - Physics: MuJoCo with two SIMPLE G1 whole-body controllers (AMO lower body, position-controlled arms and hands) stepped together. A scripted expert makes every motion of both robots; each episode's labels are SIMPLE's control targets of that robot. During noisy steps the perturbed arm targets drive the simulation only within the physics substeps and the clean targets are restored before recording, and AMO's recorded command is the clean one, so the labels are the expert's. - Boxes: 0.90 to 1.10 m long, 0.26 to 0.30 m wide, 0.20 to 0.26 m tall, 2.0 to 4.0 kg, carried 1.0 to 1.3 m sideways between two tables of equal height. - Recording: 967 runs with seeds 7001 to 8008 (seeds are not shared with the held-out evaluation seeds); 618 complete (64 %). On the calibration seeds the clean expert completes 17/20 and this perturbation level 15/20. - Rendering: MuJoCo replay of the recorded states from each robot's head camera and from the side camera, without domain randomization. - Curation: only complete runs, in which the box is released upright on the second table within 15 cm of its target (4.1 cm on average) and all hands are off it. - Checks: labels against the recorded states; for every noisy step, `action` against the clean AMO command and `action_executed` against the executed command and arm offset in the source recording (no mismatch in any episode); parquet rows against the frames of every video; loading with LeRobot 0.3.3 as Ψ₀ does. Every file was hash-checked after each copy. Recoveries in the 618 complete runs, per run (counted from the recordings): | | Per run | |---|---| | A robot more than 5 cm off its stance line / brought back within 3 cm | 5.8 / 4.4 | | Re-lifts of a sagging box | 5.8 | | A robot losing its share of the load / regaining it | 2.5 / 2.4 | | Runs with a lift under-shoot, all completed by the verified hold | 104 | The unperturbed expert already shows many of these corrections (about 6 re-lifts and 3 to 4 stance-line returns per run), so the perturbations add to its natural variation rather than create corrections it never makes. ## Known limitations - MuJoCo rendering: plain materials, no lighting or texture randomization. - The G1 wrist pitch and yaw motors are limited to 5 N·m, which keeps the grip near 35 N per hand; the box slips slightly while walking and the experts lift it back together. - No hand noise in the grasp or the set-down (the grip does not tolerate it), so those phases are perturbed only through the AMO height. - Both robots follow one script; each episode contains only one robot's observations and actions. - Ψ₀'s SIMPLE fine-tuning transform reads `observation.images.egocentric` only and ignores `action_executed` and `observation.dart_active`; training on both cameras needs a change on the Ψ₀ side. ## Files - `data/`, `meta/`, `videos/`: the LeRobot v2.1 dataset. `meta/episodes.jsonl` holds each episode's `environment_config` (role, pair, side camera, box size and mass, target position, placement result, seed, renderer, DART configuration). - `meta/stats_psi0.json`: the normalization statistics Ψ₀ uses, identical to `meta/stats.json`. - `meta/modality_head_only.json`: the modality file with the head camera alone. - `previews/`: a preview of one run. ## Loading ```python from lerobot.datasets.lerobot_dataset import LeRobotDataset ds = LeRobotDataset("randing2000/SimFactoryBoxCarryDualDART-v1") ```