Litian2002/zerograv-distill-r2-regular-gravity-rendered

Rendered regular-gravity expert trajectories, round-two distillation training set (LeRobot v2.1)

Camera-image version of episodes 0-99 of every task in stray-light/zerograv-manipulation-regular-gravity-trajectories-state-v1, rendered so that a vision-language-action model can be trained on them. This is, unchanged, the dataset on which the round-two regular-gravity-trained model of our zero-gravity distillation study was fine-tuned (pi0.5-DROID, openpi LoRA, 42,000 steps, global batch 32). The zero-gravity counterpart, rendered the same way, is Litian2002/zerograv-distill-r2-zero-gravity-rendered.

Contents

800 episodes (8 tasks x 100) and 216,888 frames at 15 fps. One parquet row is one control step.

Task	Prompt	Environment	Source episodes	Episodes	Frames
PushCube	push the cube to the target region	PushCube-DROID-Vanilla-Matched-v1	0-99	100	14,941
PickCube	pick up the cube and hold it at the target location	PickCube-DROID-Vanilla-Matched-v1	0-99	100	44,343
LiftPeg	grasp the peg and lift it upright	LiftPeg-DROID-Vanilla-Matched-v1	0-99	100	26,628
PickSingleYCB	pick up the object and hold it at the target location	PickSingleYCB-DROID-Vanilla-Matched-v1	0-99	100	46,278
PullCube	Pull the cube to the target position	PullCube-DROID-Vanilla-Matched-v1	0-99	100	18,930
PlaceSphere	place the sphere into the bin	PlaceSphere-DROID-Vanilla-Matched-WallRelease-v2	0-99	100	22,335
RollBall	roll the ball to the goal region	RollBall-DROID-Vanilla-Matched-SpawnBox-Release-Corridor-Still-w3-v1	0-99	100	20,908
PushT	push the T-shaped block onto the T-shaped target	PushT-DROID-Vanilla-Matched-v3	0-99	100	22,525

Feature	Shape	Meaning
exterior_image_1_left	180 x 320 x 3	exterior camera, placed to match the lab's real ZED 2i (intrinsics read from real footage, extrinsics solved against real images)
wrist_image_left	180 x 320 x 3	wrist camera
joint_position	7	arm joint angles, rad
gripper_position	1	0 = open ... 1 = closed
actions	8	[0:7] normalised joint-target increments, 1.0 = 0.2 rad per 15 Hz step (DROID "joint velocity"); [7] gripper command, 0 = open ... 1 = close
timestamp, frame_index, episode_index, index, task_index		LeRobot bookkeeping

The conventions are DROID's, so openpi's DroidInputs reads the data without a remap. Images are stored inside the parquet files (no videos). The layout is LeRobot codebase v2.1, the version that openpi's pinned lerobot reads; it is not the v3.0 layout of the source dataset.

episode_sources.csv has one row per episode: LeRobot episode_index, task, prompt, environment id, source dataset, source episode number, length, whether the goal is visible, and flag (see Filtering).

How it was rendered

• Each recorded simulator state in the source trajectory.h5 was written back into the simulator (no physics stepped) and both cameras were rendered, so robot (Franka + Robotiq 2F-85, the DROID configuration), objects and poses are exactly the expert's. Frame 0 is expert step 0.
• Scene ucr_droid: table and walls textured to look like the lab's real setup, instead of ManiSkill's stock scene. Rasterised (not ray-traced), stored at 320 x 180.
• Goals are visible. In PickCube and PickSingleYCB the goal is a green sphere that the source environments hide; it is drawn here (these two tasks were re-rendered on 2026-10-09 for that).
• The containment walls of the source environments have no visual geometry and do not appear in the images.
• Privileged simulator state (object pose, goal position, contact flags) is not included.

Filtering

No filtering beyond the source dataset's own collection QA: success-only episodes, per-task checks (success on the last step, a 15-frame hold, grasp held at success, arm still after the strike, goal coverage), and trimming to 15 frames after success. See the source card. Episodes 0-99 of each task are used exactly as published. Note that the two source collections used different object-ejection thresholds (1.0 m/s in zero gravity, 4.0 m/s in regular gravity; explained in the regular-gravity source card), and that the source cards' known issues carry over.

33 episodes succeed without the robot touching the object: PushCube 22/100 and PullCube 11/100. The cube spawns in the air, falls (restitution 0.9 in these tasks, see the source card's known issues) and bounces into the goal by itself. The success condition becomes true at step 2-9, and replaying the recorded states shows the gripper never closer than 178 mm to the cube surface during the whole episode (165 of the 167 other episodes of these two tasks come within 12 mm). The source QA has no robot-contact requirement, so they passed it. They are kept here because the model was trained on them; they are marked no_robot_contact in episode_sources.csv. Drop them if you need robot-caused successes only. Every other episode in this set is at least 111 frames long.

License and attribution

CC BY 4.0, as the source dataset. Expert policies, environments, trajectories and their QA: stray-light/zerograv-manipulation-regular-gravity-trajectories-state-v1. Rendering and conversion: distillation/render.py and distillation/to_lerobot.py in the project repository.