Huggingzhu1/stretch_pickup_nav_v1

Stretch 3 pickup and navigation demonstrations (MagicSim, LeRobot v2.1)

1000 expert demonstrations of a Hello Robot Stretch 3 in Isaac Sim (MagicSim), recorded for fine-tuning pi0.5 (openpi). Two language-conditioned tasks, camera-only policy inputs, every episode graded and accepted.

Split	Episodes	Frames	Purpose
train/	800	420,641	fine-tuning
validation/	106	55,345	offline checkpoint selection
test/	94	49,229	closed-loop evaluation in the simulator

Scene families (one desk layout and lighting setup with several reseeded episodes) never cross splits, so validation and test scenes are new to a model trained on train/. 10 frames per second, about 60 s per episode.

Tasks and instructions

Task	Episodes	Instruction
Pickup	789	Pick up the <noun> on the desk and hold it after retracting.
Navigation	211	Drive to the red marker and park.

Nouns: apple, ball, can, bottle, medicine bottle. People in the scene are never mentioned. Pickup episodes by scene group:

Group	Scenes	Episodes
A	apple only, no people	321
B	apple, 1 to 2 animated people walking or idling in the room	130
C	10 object types with random colours, no people	52
D	0 to 2 people plus the 10 object types with colours	286

Objects: apple_010 (484), medicine_bottle_038 (40), bottle_2 (38), apple_001 (35), medicine_bottle_004 (34), apple_006 (33), can_2 (33), can_3 (33), ball_003 (32), ball_032 (27). People per pickup episode: 0 (478), 1 (189), 2 (122).

Every scene randomizes the desk model, desk colour, floor material, background panorama, lighting, object pose and colour, robot start pose, and (B, D) the people, their paths and animations. Per-episode values are in <split>/meta/episode_metadata.jsonl (campaign, scene family, object, colour, people count, seed, desk, floor, background, lift height achieved).

Observations and actions (contract magicsim_stretch_policy_v2)

Feature	Shape	Content
observation.images.head_cam	1280 x 720 x 3 (H x W, portrait)	head RealSense D435i, real calibration (fx = fy = 912.78, rectified, rotated 90 deg clockwise as mounted on the robot)
observation.images.wrist_cam	480 x 640 x 3	wrist RealSense D405, real calibration (fx = fy = 426.92)
observation.state	11 float32	see below
action	8 float32	the expert's command at this frame; action[i] follows observation[i]
task	string	the instruction

State: lift_m, total_extension_m, head_pan_rad, head_tilt_rad, wrist_yaw_rad, wrist_pitch_rad, wrist_roll_rad, left_finger_rad, right_finger_rad, forward_m_s, yaw_rad_s.

Action, with limits:

Index	Command	Unit	Low	High	Meaning
0	forward_m_s	m/s	-0.22	0.22	base body-forward velocity
1	yaw_rad_s	rad/s	-0.45	0.45	base yaw velocity
2	lift_m_s	m/s	-0.16	0.16	lift velocity
3	total_extension_m_s	m/s	-0.18	0.18	arm extension velocity (four telescoping joints)
4	wrist_yaw_target_rad	rad	-1.57	1.571	absolute servo target
5	head_pan_target_rad	rad	-1.571	1.5	absolute servo target
6	gripper_close	binary	0	1	both fingers
7	wrist_pitch_target_rad	rad	-1.57	0.56	absolute servo target

Head tilt is held at -0.6 rad and wrist roll at 0 in every episode. Videos are H.264 (keyframe every 2 frames for random access). Full contract, camera calibration and per-episode provenance hashes: policy_contract.json.

How the data was made

• Simulator: MagicSim on Isaac Sim 5.1, Stretch 3 URDF, PhysX. Camera models use the calibration exported from a real Stretch 3 (rs-enumerate-devices -c).
• Expert: a camera-only planner. Base navigation with DWB on a depth-camera costmap, head-camera workspace probe before grasping, cuRobo motion generation for the arm, a live grasp re-plan from the object's observed pose, 1.0 s gripper close, coupled fingers, rubber-like fingertip friction (recorded per episode in extra/episode_*.json of each split's meta/).
• Grading: pickup = object lifted at least 5 cm, retracted, and still held; navigation = base parked within 2.5 cm of the marker. No contact with people, no clearance violations. Only accepted episodes are included; roughly 13 to 15 % of attempts were rejected (grasp slips, perception stalls).
• Export: MagicSim native recordings converted to LeRobot v2.1 with scripts/hri_pi05/ from the MagicSim repository; no frames are dropped and observation i is paired with action i.

Loading

from huggingface_hub import snapshot_download
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

root = snapshot_download("<repo_id>", repo_type="dataset")
train = LeRobotDataset("train", root=f"{root}/train")       # lerobot v2.1 API
sample = train[0]
sample["observation.images.head_cam"].shape   # (3, 1280, 720)
sample["action"]                              # (8,)
sample["task"]                                # instruction string

For openpi, point HF_LEROBOT_HOME at the downloaded root and use repo_id="train"; the training configuration used for the published results is scripts/hri_pi05/hri_pi05_training.py (--recipe saif).

Baseline results

Two pi0.5 LoRA fine-tunes (LoRA on both towers, cosine 2.5e-5 to 2.5e-6 with 10 % warmup, batch 32, 10,000 steps, action horizon 10, one A100):

Model	State input	Train loss	Validation loss	Closed-loop test (sim)
hidden-state	9 of 11 (head pan, wrist yaw zeroed)	0.0041	0.0069	0 / 4
full-state	all 11	0.0037	0.0040	0 / 2

Both models reproduce the expert almost exactly when shown recorded frames (mean error 0.001 m/s on speed, 0.014 rad on head pan, 99.7 % gripper agreement) and fail in closed loop by compounding error: an early over-turn moves the camera view out of the training distribution. The dataset contains no recoveries from off-course states; that is the known gap.

Files

train/ validation/ test/      LeRobot v2.1 datasets (data/, videos/, meta/)
<split>/meta/episode_metadata.jsonl   scene metadata per episode
policy_contract.json           contract, limits, calibration, per-episode provenance
examples/                      one episode per group: head + wrist video, instruction, grader verdict, resolved scene

Citation

MagicSim, Northwestern University, 2026. Expert, simulator and export code: the MagicSim repository (src/magicsim/HRI, scripts/hri_pi05).