EmbodiedSWE/clear_organic_agent_ab_A1000.jointtarget_20hz_20260918
This dataset was created using LeRobot. About this dataset (agent-in-the-loop A/B, arm A, first 1000 verified successes) Simulated Franka Panda demonstrations of a produce-clearing task (pick every organic item - lemons, a lime, an orange, a small pumpkin - off a cluttered packing table and drop it into the crate, leaving the distractors), generated in Isaac Sim by the CoSiGen data-generation pipeline. Every episode is a certified success of the task grader (all present organics… See the full description on the dataset page: https://huggingface.co/datasets/EmbodiedSWE/clear_organic_agent_ab_A1000.jointtarget_20hz_20260918.
This dataset was created using LeRobot.
About this dataset (agent-in-the-loop A/B, arm A, first 1000 verified successes)
Simulated Franka Panda demonstrations of a produce-clearing task (pick every organic item - lemons, a lime, an orange, a small pumpkin - off a cluttered packing table and drop it into the crate, leaving the distractors), generated in Isaac Sim by the CoSiGen data-generation pipeline. Every episode is a certified success of the task grader (all present organics in the crate, no distractor in it). Grasps are real friction pinches (no weld).
- A/B arm A: NO agent: the delivered oracle re-run with scripted executed-action noise (gaussian 0.02 rad on the 7 arm targets, 0.4 s bursts at p=0.03 per 20 Hz step, only while the hand is above 0.85 m, i.e. transport phases) plus per-episode produce friction (0.90-1.30 / 0.80-1.10) and organic mass (x0.7-1.3) draws. Labels are the clean commands; the noise is executed below the recorder. Held out from every training scene: pomegranate01 and orange01 (the novel-object test items). Design and chronicle: experiments/2026-09-18clearorganicagent_ab/README.md in the CoSiGen repo.
- Episodes / frames: 1000 / 2,103,185 at 20 Hz (29.2 h)
- Cameras:
front,side,wrist- 480x640, h264 - State (8):
panda_joint1..7measured positions (rad) + gripper closedness in [0, 1] (0 = open) - Action (8): ABSOLUTE
panda_joint1..7_target(rad) latched at 20 Hz + commanded gripper closedness. - Control law: stationary - the scene's stock Franka joint PD (kp 400 / kd 80) and gripper drives; deploy through the plain 20 Hz joint-target executor (
meta/bake.json->controller). - Layout randomization: organic subset 4-5 of 5, in-group slot shuffle, +-2.5 cm jitter, free yaw; Halton draws with the solve index offset by 4093 from the physical index (independent friction and expert-parameter draws).
- Idle filter: stationary ticks dropped at bake time (0).
- Extra column:
raw_command(9) - the raw controller command vector, not a policy input.
