Autobrik/ecse552-ocbench-block

ECSE 552: OCBench single-block demonstrations

Scripted-expert demonstrations for the McGill ECSE 552 project Learning to Manipulate a New Object from Action-Free Videos (code). A UR5e arm with a Robotiq gripper (OCBench, MuJoCo) moves a cube into a goal square, seen by three RGB cameras.

Every episode stores images and actions. In the project, actions of target objects (everything except source) are hidden from every method except the true-action oracle and evaluation: those collections stand in for action-free videos.

Collections

Folder	Object	Episodes	Seeds	Expert success	Transitions	Mean length
source-n300-seed0/	source: red 6x6x6 cm cube	300	0-308	97.1%	24568	81.9
A-n200-seed100000/	A: blue 6x6x6 cm cube (appearance)	200	100000-100212	93.9%	17416	87.1

source is the labeled cube; target A changes only the cube colour. The same seed gives the same initial cube and arm state for every object.

Scene

• Environment visual-block-cpu-lite-single-task1-v0 from the OCBench fork, 10 Hz control, at most 200 steps (20 s).
• Cameras: front, side, wrist, 224 x 224 RGB, one image per control step.
• Actions: 7 normalized values in [-1, 1], six joint increments and one gripper increment.
• Goal: fixed at (0.425, 0) m, marked by a dark-gray square of 12 cm inner side. Success requires the whole cube inside the square, resting on the table, released and still for 1 s; episodes stop at success.
• Initial state: the cube in a 20 x 20 cm area beside the goal (x 0.32-0.52 m, y 0.15-0.35 m, uniform yaw); the arm always starts from the same pose (effector at (0.35, 0, 0.30) m, yaw 0). Stock OCBench samples both much more widely.
• Only successful expert episodes are kept; failed seeds are skipped (see manifest.json).

Files

Each folder holds episode_XXXXX.npz files (numbered in seed order), manifest.json (every attempt with its seed and outcome, plus the scene parameters), initial_states.csv (cube position and yaw, arm joints and split of every episode), provenance.txt (command and commits) and packages.txt. One episode with T steps:

Key	Shape	Type	Meaning
observations	(T+1, 3, 224, 224, 3)	uint8	front, side, wrist images before each step, plus the final frame
actions	(T, 7)	float32	action applied at each step (exactly the values stepped)
rewards, masks, terminals	(T,)		OCBench reward, 1 - terminated, episode end
qpos, qvel	(T, ...)	float32	simulator state before each step: privileged, never a model input
seed, success, length	scalars

observations[t + 1] is the result of actions[t] applied in state observations[t]: replaying the stored actions from the seed reproduces qpos exactly.

import numpy as np
ep = np.load("source-n300-seed0/episode_00000.npz")
front, side, wrist = ep["observations"][0]

Splits are per episode, in blocks of ten (8 train, 1 validation, 1 test), so the first 100 episodes split 80/10/10 and keep their split in the full 300 (split_of in src/action_free/data.py).

Reproduce

git clone --recurse-submodules https://github.com/BrikHMP18/ecse-552-project.git && cd ecse-552-project
git checkout 851bd06bb7018186612f256683bc42231f7a03cf && git submodule update --init --recursive
uv sync --locked
MUJOCO_GL=egl uv run python scripts/collect_demos.py --object source --num-episodes 300 --base-seed 0

Simulator state is bit-exact per seed; GPU (EGL) images can differ from the stored ones in about 0.03% of pixels (shadow edges).

License and credit

MIT, like OCBench (arXiv 2610.07056), the benchmark this data is generated with.