RoboCasa-DC

RoboCasa-DC is a demo-conditioned extension of RoboCasa for benchmarking policies that are conditioned on a single demonstration video. It pairs RoboCasa's Panda-arm trajectories with newly collected GR-1 humanoid demonstrations, which serve as a simulated proxy for human demonstrations, across 24 RoboCasa kitchen manipulation tasks.

RoboCasa-DC was introduced in SeeTraceAct: Visibility-Aware Latent Planning from Cross-Embodiment Demonstration Videos.

• Project: https://jaehyeon-son.github.io/seetraceact/
• Paper: https://arxiv.org/abs/2606.02745
• Code: https://github.com/jaehyeon-son/SeeTraceAct

Overview of RoboCasa-DC. For each of the 24 tasks, 100 Panda-arm trajectories are paired with GR-1 humanoid demonstrations for training. For evaluation, each task has 50 pre-defined seeds, each with a GR-1 humanoid demonstration.

Benchmark construction

• Training pairs. For each task, 100 Panda-arm trajectories from RoboCasa are paired with GR-1 humanoid demonstrations. Each demonstration was collected by restoring the initial simulation state of its Panda-arm trajectory and teleoperating the humanoid in the same scene with a Leap Motion controller.
• Evaluation seeds and demonstrations. Each task has 50 pre-defined evaluation seeds, each with an environment in eval_envs/ and a GR-1 humanoid demonstration collected from the same initial scene state.
• Demonstration camera view. For each task, the GR-1 camera view that best captures task execution was selected manually; it is the single camera listed in the task's meta/info.json.

Contents

directory	contents	episodes
panda_train/<Task>/	Panda-arm trajectories from RoboCasa: actions, proprioceptive states, videos from two static cameras and one wrist camera, end-effector traces	24 tasks x 100 = 2400
gr1_train/<Task>/	GR-1 humanoid demonstrations paired with panda_train: one camera video, state/action	2369
gr1_eval/<Task>/	GR-1 humanoid demonstrations for the evaluation seeds: one camera video, state/action	24 tasks x 50 = 1200
eval_envs/<Task>/<seed>/	Panda-arm simulation scene for each pre-defined evaluation seed	24 tasks x 50 = 1200

Pairing rule (by episode index)

Pairing is purely positional; there is no separate index file.

• Training: panda_train/<Task>/.../episode_XXXXXX and gr1_train/<Task>/.../episode_XXXXXX with the same task and the same 6-digit episode index form one training pair.
• Evaluation: seed directory eval_envs/<Task>/<N>/ pairs with gr1_eval/<Task>/.../episode_{N:06d}, which shares seed N's initial scene state. The policy is conditioned on the GR-1 video of episode N and rolled out in the Panda-arm scene of seed N. Each task has exactly 50 pre-defined seeds. The gr1_eval episode indices match the seed directories exactly.

File formats

Episode files follow the GR00T / LeRobot v2 layout: one parquet and one mp4 per episode per camera, data/chunk-000/episode_XXXXXX.parquet and videos/chunk-000/observation.images.<camera_key>/episode_XXXXXX.mp4. All videos are 20 fps; Panda videos are 128x128, GR-1 videos are 256x256. Frame t of the video matches row t of the data and trace parquet.

panda_train/<Task>/

• meta/modality.json defines the named slices of the state and action vectors; meta/tasks.jsonl maps task_index to the language instruction.

• data/chunk-000/episode_XXXXXX.parquet, one row per timestep:

  column	type	description
  observation.state	list[53]	joint pos / cos / sin / vel, eef pos + quat, gripper qpos / qvel, base pos + quat, base-to-eef pos + quat (slices in meta/modality.json)
  action	list[12]	delta eef position (3), delta eef rotation as axis-angle (3), gripper (1), mobile base (3), torso (1), base mode (1)
  annotation.human.action.task_description	int64	index into meta/tasks.jsonl (language instruction)
  annotation.human.validity	int64	index into meta/tasks.jsonl

• trace/episode_XXXXXX.parquet: 2D end-effector traces, generated by projecting the end-effector position into each of the two static camera views. Columns are a two-level pandas MultiIndex (read with pandas.read_parquet): ('eef_trace_robot0_agentview_left_pixel_rc', 'x' | 'y') and ('eef_trace_robot0_agentview_right_pixel_rc', 'x' | 'y'), float32 image coordinates normalized by image size. Values outside [0, 1] mean the end effector is off-screen in that view.

gr1_train/<Task>/ and gr1_eval/<Task>/

• meta/info.json (paths, video shape, fps). gr1_train also has meta/camera_config.json, the camera pose used to render the GR-1 videos.
• data/chunk-000/episode_XXXXXX.parquet: same columns as Panda, with observation.state of length 221 and action of length 24 (GR-1 humanoid), plus an int64 column observation.images.<camera_key> holding the frame index into the mp4. The GR-1 splits have no meta/tasks.jsonl, so their annotation columns carry no language; the instruction for a training pair is the Panda episode's.

eval_envs/<Task>/<seed>/

• environment.xml: MuJoCo model of the Panda kitchen scene.
• ep_meta.json: RoboCasa episode metadata (layout / style ids, object and fixture configs, language instruction, camera configs).
• initial_state.npz: states, the flattened MuJoCo simulation state to reset to.

License

RoboCasa-DC is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. It is derived from RoboCasa, whose assets and datasets are also licensed under CC BY 4.0. The SeeTraceAct code is released separately on GitHub.

Citation

If you use RoboCasa-DC, please cite:

@inproceedings{son2026seetraceact,
  title     = {SeeTraceAct: Visibility-Aware Latent Planning from Cross-Embodiment Demonstration Videos},
  author    = {Son, Jaehyeon and Kim, Junhyun and Kam, Kyle and Coholich, Jeremiah and Kim, Seok Joon and Kim, Jinhoo and Kim, Chris Dongjoo and Cho, Jaemin and Fox, Dieter and Kira, Zsolt},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}