NoahWeiss/farm_uf850

FARM UF850 — Pick-and-Place with a UFactory UF850

LeRobot-format dataset of teleoperated single-arm manipulation collected on a UFactory UF850 (6-DoF arm with parallel gripper) using a Meta Quest as the teleop controller. 66 episodes spanning 17 distinct natural- language tasks, primarily cube manipulation (pick the cube and place on a box) including occlusion and failure-recovery scenarios.

Collected with the FARM edge agent (repository) as a CS153 final project.

Quick stats

Episodes	66
Total frames	25,803
Duration	14.3 min
Episode length (median)	12 s
FPS	30
Robot	UFactory UF850, 6-DoF + parallel gripper (7-D state/action)
Cameras	2 × Intel RealSense D435 (base + wrist), 640×480, h264
Unique tasks	17
Format	LeRobot v2.0

State / action space

7-dim float32 in both observation.state and action:

idx	name	unit	range
0	joint1	rad	~[-0.94, 0.83]
1	joint2	rad	~[-1.45, -0.07]
2	joint3	rad	~[-1.88, -0.18]
3	joint4	rad	~[-2.71, 2.22]
4	joint5	rad	~[ 0.26, 1.82]
5	joint6	rad	~[-1.18, 0.78]
6	gripper	0 (open) → 1 (closed)	observed 0.0–0.39

action[t] = observation.state[t+1]; the final frame's action repeats the final state so chunks line up.

Tasks

All descriptions are imperative, lowercase, no terminal punctuation.

idx	n_episodes	task
0	4	pick up the rubiks cube from the yellow side and place it on the box
1	3	pick up the rubiks cube from the white side and place it on the box
2	1	pick up the rubiks cube from the yellow side, drop it, then pick it up and place it on the box
3	1	pick up the rubiks cube from the right side and place it on the box
4	30	pick up the cube and place it on the box
5	1	see the cube behind the box, push the box sideways, then pick up the cube and place it on the box
6	1	pick up the cube from the right side and place it on the box
7	2	pick up the cube and place it on the box but it falls off, then pick it up and place it on the box
8	5	see the cube behind the box, pick up the box and move it aside, then pick up the cube and place it on the box
9	1	pick up the cube and place it on the box, knock it off while moving away, then pick it up and place it on the box
10	3	pick up the cube, drop it while placing it on the box, then pick it up and place it on the box
11	1	pick up the cube and misplace it so it falls off, then pick it up and place it properly on the box
12	1	pick up the cube from behind the box and place it on the box
13	1	pick up the cube, drop it by accident, then pick it back up and place it on the box
14	5	pick up the box, see no cube under it, then place the box back down
15	3	pick up the box and place it back down
16	3	pick up the box, find a cube under it, place the box aside, pick up the cube, then place it on the box

Layout

farm_uf850/
├── meta/
│   ├── info.json         # LeRobot v2.0 manifest
│   ├── episodes.jsonl    # one row per episode {episode_index, tasks, length}
│   ├── tasks.jsonl       # one row per task {task_index, task}
│   └── stats.json        # per-feature mean/std/min/max
├── data/chunk-000/
│   └── episode_NNNNNN.parquet  ×66
└── videos/chunk-000/
    ├── observation.images.base/    # h264 mp4 ×66
    └── observation.images.wrist/   # h264 mp4 ×66

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("YOUR_USERNAME/farm_uf850")   # streams from the Hub
print(ds.num_episodes, ds.num_frames, ds.fps)     # 66, 25803, 30
sample = ds[0]
sample["observation.state"].shape                  # torch.Size([7])
sample["observation.images.base"].shape            # torch.Size([3, 480, 640])
sample["task"]                                     # natural-language task string

Collection notes

• Hardware: UF850, 2× D435 (base + wrist), Meta Quest 2 controller, Mac mini host.
• Teleop: VR controller pose → Quest publishes ROS-TCP → laptop bridge → IK → UF850 joint commands. Per-joint One-Euro filter on the joint stream.
• Data captured at 30 Hz with motion-scaled controller input (motion_scale configurable 0.2×–4× per session).
• Frames where the cam subprocess returned a duplicate (cached) JPEG have been excluded by re-recording — the 41 demos collected before the camera FPS fix are not in this release.
• Saved data reflects the smoothed physical robot motion, not the unsmoothed digital-arm target (encoder reads from the UF850 SDK at 250 Hz, snapshotted into frames at 30 Hz).