Datasets:

Numbnut
/

stackblocks-stage3-v1

license: mit
task_categories:
  - robotics
tags:
  - LeRobot
  - robotwin
  - sim
  - franka
configs:
  - config_name: default
    data_files: data/*/*.parquet

stackblocks-stage3-v1 — RoboTwin stack-blocks-two task finetune data

Stage-3 of a 3-stage curriculum (DROID CoT pretrain → playground @grip conditioning → task finetune). 45 human-teleop episodes of the stack_blocks_two task (Franka, RoboTwin sim) — the source dataset of the T-series / EE-noCoT ablation campaigns. LeRobot v2.1 video/h264, schema-identical to Numbnut/playground-stage2-v1 and Numbnut/droid-keeps-v6:

observation.images.ext = bird_eye (the <loc>/@grip grounding view), 320×240
observation.images.wrist = right_camera
observation.state (8,) = joint qpos (7) + gripper open (1=open)
action (7,) = EE delta-euler + gripper (compute_ee_delta_actions)
task = the REAL per-episode instruction (goal-directed, e.g. "Relocate red block to the middle and place green block on it.")

Conversion specifics (see droid_provenance.json):

30→15fps stride 2 — tempo parity with stage-2. NOTE: the historical T-series/EE repos were native-rate; per-stored-frame deltas here are ~2× those — do not compare models trained here against the historical EE-noCoT/T-series numbers directly.
Static-prefix trim 0.05 rad (T-series-proven; mean ~14 frames): trim maps each episode's source-frame offset. Any future CoT span table must map kf_lerobot = (kf_source - trim) // stride, dropping kf inside the trimmed prefix.
held_out = []: stage-3 eval is closed-loop sim (run_benchmark.py --episodes 50 --seed-offset 100 --step-limit 800), all 45 episodes train. Keep cot_exclude_held_out=False for this repo.

No CoT span table yet: the existing annotation_v5.json files in the source tree are an older API format (per-kf A only); v5-span regeneration (template annotator) is a pending, separately-reviewed step.

Built by policy/pi05-pure/scripts/convert_franka50_to_lerobot.py in the RoboTwin fork.