physicalai-bmi/forge-arm-pixels

Real MuJoCo pixels captured live from the Institute's in-browser Forge arm (WebGPU), paired with the action the released state-checkpoint took. This is the exact training set behind physicalai-bmi/nano-vla-pixels.

• 2,500 frames across 128 reaches, frames/f#####.png (the rendered MuJoCo arm, 844×520).
• meta.json — per-frame { i, act:[3], obs:[7], reaches }; act is the 3-D joint-delta action, reaches is the episode index (use it for an episode-level split).

How it was made

Captured in a headless browser: the state checkpoint forge-arm-reach-bc drove the real Forge MuJoCo arm on WebGPU, and each frame was grabbed by screenshotting the WebGPU canvas (a page's own JS can't read WebGPU pixels; an external compositor screenshot can), synced to the policy's action. Capture harness: capture.cjs in the Institute repo. CC-BY-4.0.

Load (Python)

import json, glob, cv2, numpy as np
meta = json.load(open("meta.json")); frames = sorted(glob.glob("frames/f*.png"))
X = np.stack([cv2.resize(cv2.imread(f), (48,48)) for f in frames])   # or full-res
Y = np.array([m["act"] for m in meta]); ep = np.array([m["reaches"] for m in meta])

Stacking 3 consecutive frames (velocity) lifts a pixel policy from 46.6% → 82.9% held-out variance explained — see the model card.