amiryanj/so101-sim-pick — dataset card

Source of truth for the card published at https://huggingface.co/datasets/amiryanj/so101-sim-pick.

Kept here because v1's card lived only on the Hub: there was no copy to edit, diff or review. Push it with vision/push_dataset.py (or by hand) rather than editing the web form, so the repo and the Hub cannot drift.

---

v1.1 — 120 episodes of pick-and-place in MuJoCo: 100 successes and 20 failures, 34,680 frames at 30 fps.

An SO-101 follower arm picks a 40 mm basketball off a white plate and places it to one side. No human teleoperation and no real hardware: every episode is produced by a scripted state machine, so the dataset is cheap to regenerate and to grow.

• Homepage: https://github.com/amiryanj/robot_bazoo
• Simulator: MuJoCo 3.8.1, scene built by mujoco_lab/scene.py from config/sim_scene.json
• License: apache-2.0

What changed from v1

v1 was a self-consistent sim dataset — it was simply not built from the real rig's geometry, because that had not been measured yet. v1.1 is.

	v1	v1.1
episodes	82, successes only	120, successes and failures
external camera	invented pose, fovy 45	measured pose, fovy 64.7
wrist camera	guessed pose, fovy 80	mount-derived pose, fovy 40.4
plate	200 mm slab, full XML extent	12.6 mm, ending 10 mm past the last marker
background	MuJoCo's blue checkerboard	wooden desk, colour solved to match the real one
markers	none	the real tag 6 and checkerboard, as textures
ball	plain orange sphere	basketball texture
task strings	one, constant	two — success and failure

How the camera geometry was measured, and the traps involved, is written up in REAL2SIM.md.

Failures are included — read this before training

20 of the 120 episodes (17%) are failed attempts, and they are kept on purpose. They are the only examples in the dataset of what going wrong looks like, and a policy that has never seen one cannot be asked to recover. They are marked two ways:

# a float feature, 1.0 = success, 0.0 = failure
frame["success"]

# and the task sentence
"pick up the ball and place it to the side"                    # success
"pick up the ball and place it to the side (failed attempt)"   # failure

Training on everything indiscriminately will teach the policy to reproduce failures too. Filter unless you specifically want them:

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("amiryanj/so101-sim-pick", video_backend="pyav")

video_backend="pyav" is worth passing: torchcodec fails to load against some torch builds, and the error it raises does not mention the dataset at all.

How the episodes were made

A deterministic state machine, not a neural policy and not a human:

1. the ball pose is read from the simulator (no perception in the loop),
2. damped-least-squares IK plans home → high → standoff → grasp,
3. the last leg runs as a straight Cartesian line — a joint-space swing knocks the ball 20–40 mm, per-step IK nudges it 0.2 mm,
4. the jaws close, the arm lifts 10 cm, carries, and releases.

Each episode starts at the settle at HOME, so the whole reach is in the data, not just the grasp.

One honest caveat: the grasp uses a weld

The SO-101 jaws form a shallow V. Measured over the contact set, the best opposing-normal dot product is −0.49, never near −1, so the fingers squeeze a sphere sideways and it rolls free. Rigid-body contact alone cannot hold this ball. A MuJoCo weld equality is therefore switched on when both fingers touch and released when the jaws are commanded open.

Everything else — arm dynamics, contacts, the ball's motion before and after the grasp — is ordinary physics. But a policy trained here learns when to close, not how to maintain a real friction grip. This has not yet been tested against the real gripper. Expect that gap when transferring to hardware.

Cameras

key	view	resolution	fps
observation.images.external	fixed, looking at the workspace	480×640	30
observation.images.wrist	on the printed mount, above the jaws	480×640	30

The external camera sits at its measured pose and is verified two independent ways: rendering the sim at measured joint angles puts tag 6, the checkerboard and the gripper within ~10 px of a real Realsense frame, and running the real ArUco and checkerboard detectors on a sim render finds them within 0.77 px and 0.70 px of the projected measured geometry.

The wrist camera pose is a hand fit, not a measurement. Its lens was calibrated (0.656 px), but where it sits on the gripper has never been checked against a real wrist image. Treat the wrist stream as approximate.

Randomisation

Per episode: ball position, ball size and brightness, plate tint, light position and headlight intensity, start pose, approach angle, and small jitter on both camera poses and fields of view.

Not randomised: the background, the desk colour, and the overall lighting character — all three are pinned to one real lighting condition. That is the main appearance gap left.

Units

Joint positions and actions are in degrees, in the order shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper.

Two conventions to know before comparing against a real SO-101:

• the sim's wrist_roll zero is −90° from the real robot's,
• the gripper is logged in jaw degrees here, while lerobot uses 0–100 on hardware. The ranges happen to overlap; they are not the same unit.