Perceptron open-sources EgoHome: 137 egocentric clips annotated down to 100ms
rohanpaul_ai · x · 2026-11-11
Alongside Perceptron Egocentric 1.0, the team released EgoHome, the first egocentric benchmark with four annotation types per video and segments as short as 100ms.
Key points:
- 137 densely human-annotated egocentric clips of household tasks, 2 hours total.
- Four annotation categories: hand bounding boxes, subtask annotations, left/right hand actions (Ego4D closed-vocab schema), and left/right hand captions; raw annotations plus evaluation-metric code are open-sourced, with metrics designed to capture the subjectivity of open-vocabulary video description.
- Motivation: recent work from Physical Intelligence, Generalist, and Dyna shows diverse pretraining and high-fidelity physical-interaction data drive reliable control; dense egocentric annotation is prized in robotics for scalability.
- The attached leaderboard shows GPT-6 Astra (4 fps) leading hand-action scoring at 0.498, ahead of Perceptron's own model (0.421) and Gemini 3.8 Flash (0.405).
More from Embodied
- Russ Tedrake's CoRL talk: deploy robots early, train on intervention data to scale — DominiqueCAPaul · 2026-11-11
- La-JEPA: Huawei-led 1B robot policy beats 12.4B Fast-WAM (60.8% vs 39.9%) and runs 5x faster — burny_tech · 2026-11-11
- Gemini Robotics humanoid serves snacks live on CoRL 2025 show floor — ki_ki_ki1 · 2026-11-11
- SuperMimic open-sourced at CoRL: mimics 129K motions with 1000x fewer GPU-hours than SOTA — kevin_zakka · 2026-11-11
- PolicyIK replaces numerical inverse kinematics with an RL policy, beating 4 classic IK controllers on a real robot — rsasaki0109 · 2026-11-11
- MIT's FLEX pipeline lets robots refine task representations from natural language corrections — andreea7b · 2026-11-11