ZihuiGao/dpsk-robot-policy

π₀.₅ + DeepSeek-Flash hybrid rollouts on RoboDojo

28 complete bimanual manipulation episodes from a hybrid policy in which π₀.₅ generates actions and deepseek-flash reviews them, run on the RoboDojo simulation benchmark.

This is a partial replication of GPT 6 Astra as an Embodied Policy with only the reviewing model changed. The architecture, tasks, evaluation panel, seeds and scoring are the original authors'. At each decision point π₀.₅ proposes a 50-step joint-space chunk and the reviewer either accepts a 1–15 step prefix or issues a 1–5 step end-effector correction.

Contents

Episodes	28
Successful	4
Control steps	22,842
Corrected steps	1,964 (8.6 %)
Robot	dual ARX X5, 14-DoF (6+1 per arm)
Cameras	cam_high, cam_left_wrist, cam_right_wrist, 640×480
Control rate	25 Hz (control_dt 0.04 s)

Ten RoboDojo tasks: organize_table, classify_objects_by_language, imitate_sorting_sequence, arrange_largest_number, pack_objects_into_box, classify_objects, build_tower, make_kong, fold_clothes, put_bottles_into_dustbin.

Layout

metadata.jsonl              one row per episode: seeds, outcome, step counts
episodes/<case_id>/
  sim/
    <episode_id>/
      NNNNNN.npz            per-step simulator state snapshot
      action_NNNNNN.json    per-step executed 14-D action
      codex_decisions/      the reviewer's reasoning at each decision
    sensors.mp4             three camera views
    evaluation_outcome.json RoboDojo's native success and score
    scene_layout.json       initial object placement
    resolved_config.json    the fully resolved simulator/robot/camera config
  controller/
    history.json            decision-level trajectory
    progress.json           student / edited / recovery step counts
    execution_NNN.npz       actions actually executed per segment
    proposals/*.npz         π₀.₅'s raw 50-step proposals
    observations/*.png      camera frames at each decision point
    codex_workspace/        the reviewer's own scratch space and RPC records
    debug_video/            rollout video with decision annotations
    token_usage.json
  evaluation_case.json      the frozen case identity and its five seeds
  logs/                     simulator, policy and controller logs
source_snapshot/            the pipeline code as run (identical for all episodes)

action_NNNNNN.json holds executed_action, a 14-vector ordered left_joints[6], left_gripper, right_joints[6], right_gripper, with the gripper continuous in [0, 1] where 0 is closed.

Comparing controller/proposals/*.npz against controller/execution_*.npz shows what π₀.₅ suggested versus what actually ran — the difference is exactly what the reviewer changed.

Loading

from datasets import load_dataset
meta = load_dataset("json", data_files="metadata.jsonl")["train"]

metadata.jsonl is the index; the per-step arrays live in the episode folders and are read directly with numpy.load / json.

How it was produced

RoboDojo at commit ee67a14 on Isaac Sim 5.1, π₀.₅ served from the RoboDojo-finetuned checkpoint RoboDojo-sim-arx_x5-joint-0/59999 through OpenPI/JAX, reviewer deepseek-flash at max reasoning effort over the Responses API. Hardware was a single NVIDIA H200; four episodes ran concurrently.

Limitations

• 28 of the panel's 50 cases. The run stopped when the API account ran out of credit, so the missing 22 are the ones in flight at that moment — not a random sample.
• Different hardware from the original report (H200/driver 580 against RTX 4090/driver 570). RoboDojo's documentation notes run-to-run variation across hardware, drivers and timing, and it was observed directly here: one case with identical seeds scored 1.00 in one run and 0.10 in another. Treat single-episode differences as noise.
• max in place of xhigh. DeepSeek exposes low/high/max and has no xhigh, so its top tier was substituted for the report's setting.
• Four concurrent episodes, where the report ran one container per case. Wall-clock latency in these logs is therefore inflated; a clean single-episode measurement gave a median of 27.1 s per decision and a 48× wall-to-physical ratio.
• Success here (4/28) is well below the original report's rate on the same cases. That gap has not been isolated to the model alone.

Credit and licensing

The benchmark, tasks, evaluation panel, seeds, scoring and the hybrid architecture are the work of the RoboDojo authors and of the original report's authors. This dataset contains only rollouts produced by running their pipeline with a different reviewing model.

RoboDojo is released under a non-commercial research license, which governs the simulator assets and task definitions these trajectories were generated from. No simulator assets and no model checkpoints are redistributed here.

source_snapshot/ is the original report's pipeline code exactly as it ran, including its prompt files (SKILL.md, gate_prompt.md, teacher_context.md, eef_control.md). That code is MIT-licensed by its authors and is kept here so these rollouts can be audited against the precise instructions that produced them; credit for it belongs to them, not to this dataset.