Boyun7/GroundProbe-dataset

GroundProbe Teleoperation Demonstrations

Human teleoperation demonstrations for GroundProbe, a diagnostic benchmark for vision-language-action (VLA) manipulation policies. A Franka Panda in NVIDIA Isaac Lab-Arena was driven with a Meta Quest 2 controller; every episode here ended in task success.

• 540 episodes, 51,234 frames across 12 sub-datasets
• LeRobot v2.0 layout: parquet per episode, one mp4 per camera per episode
• Two 640×640 RGB cameras (third-person and wrist)
• Three parallel action encodings per step (see below)
• Code, scenes and evaluation harness: https://github.com/AndersonYu7/Benchmark

Sub-datasets

Each top-level folder is an independent LeRobot dataset with its own meta/, data/ and videos/.

folder	axis	episodes	frames	instructions
quest_l2_cubes_pilot	L2 spatial modifier — cubes	241	22,434	5
quest_l2_fruits_pilot	L2 spatial modifier — fruits	47	3,924	1
quest_l2_tools_pilot	L2 spatial modifier — tools	21	1,964	1
quest_l3_cubes_pilot	L3 negation — cubes	74	6,930	5
quest_l3_fruits_pilot	L3 negation — fruits	32	3,050	5
quest_l4_basic_pilot	L4 relative size — basic	27	2,080	3
quest_l4_cluttered_pilot	L4 relative size — cluttered	26	2,186	3
quest_l4_dense_pilot	L4 relative size — dense	26	2,121	3
quest_d2_pilot	D2 compositional pairing	17	1,491	12
quest_d3_v1_pilot	D3 state-conditioned — V1	10	907	1
quest_d3_v2_pilot	D3 state-conditioned — V2	9	1,623	1
quest_d3_v3_pilot	D3 state-conditioned — V3	10	2,524	1

These are pilot collections; coverage per (instruction, complexity) cell is uneven and some cells hold a single episode.

Download

pip install -U huggingface_hub
hf download Boyun7/GroundProbe-dataset --repo-type dataset --local-dir dataset

A single sub-dataset:

hf download Boyun7/GroundProbe-dataset --repo-type dataset \
    --include "quest_l2_cubes_pilot/*" --local-dir dataset

Features

key	dtype	shape	meaning
observation.images.third_camera	video	640×640×3	fixed third-person view
observation.images.wrist_camera	video	640×640×3	camera on panda_hand
observation.state	float32	9	7 arm joint positions + 2 finger positions
observation.tcp_pose	float32	7	end-effector pose in world frame, [x, y, z, qw, qx, qy, qz]
action	float32	7	aim-point offset, [dx, dy, dz, droll, dpitch, dyaw, gripper]
action.tcp_disp	float32	7	displacement to the next frame, same layout
action.joint	float32	8	absolute joint targets + gripper
next.done, next.success	bool	1	episode termination / success

Rotations are XYZ-extrinsic Euler angles in radians, world frame, post-multiplied (R_target = R_delta · R_current). Gripper is ±1.

Per-episode metadata in meta/episodes.jsonl carries groundlab.* fields: instruction_id, complexity (clean / cluttered / multi), layout_seed (reproduces the object arrangement), target_asset, and split.

The three action encodings are not interchangeable

They describe different quantities, and picking the wrong one for a given controller is the most common way to get a policy that "almost works".

• action is the IK target the operator's controller commanded at that step, minus the measured TCP pose. It is an aim point, not a motion: the simulated differential-IK controller closes only about 19% of a commanded gap per recorded step (median 55 mm commanded, 11 mm moved). Decode it by adding it back to the TCP pose it was paired with. This is the only Cartesian encoding that replays open-loop exactly.
• action.tcp_disp is tcp_pose[k+1] − tcp_pose[k]: what the arm actually did. It is the convention most VLAs use, but under this controller commanding a recorded displacement moves the arm about a fifth as far. The last frame of every episode is zero.
• action.joint is the joint-position target the IK term issued. It is absolute, so a whole action chunk can be executed without re-observing.

observation.tcp_pose is the ee_frame "end_effector" frame at panda_hand + 0.1034 m along the tool axis, while the IK controller drives a point at panda_hand + 0.107 m. The 3.6 mm offset cancels for action but biases action.tcp_disp if a displacement is sent to the controller unchanged.

In ACT experiments on quest_l2_cubes_pilot (instruction T1, clean), the same checkpoint architecture reached 85% with action, 80% with action.joint, and at best 50% with action.tcp_disp on training layouts.

Control rate

info.json records fps: 31 because LeRobot stores an integer. The true recording rate is 31.25 Hz (sim_dt 0.004 s × decimation 4 × 2 env steps per recorded frame), given exactly in the control block of each meta/info.json. Use that value when reproducing the cadence.

Known issues

• Trimmed episode starts in quest_l2_cubes_pilot. An earlier post-processing pass removed leading frames until the TCP had moved 3 cm, so roughly 76 of its 241 episodes (the low indices, below 90) begin mid-reach and already in motion rather than at rest at the initial pose. Episodes at index 90 and above start at rest. No other sub-dataset is affected. Rollouts that start from rest will look out-of-distribution to a policy trained heavily on the trimmed cells (notably T3).
• Uneven coverage. These are pilot collections; several cells have one or two episodes.
• Layout reproduction needs the simulator. layout_seed reproduces the arrangement only through the scene code in the GitHub repository, with the same obstacle set and random-number consumption order as the collector.

License

Apache 2.0. Scene and object assets used to render these episodes come from upstream projects under their own licenses; see the GitHub repository's assets/README.md.