mjfit peg benchmark under fire: cylinder-peg setup embeds a strong prior
YouJiacheng · x · 2026-10-20
In a discussion with Younghyo Park, YouJiacheng flags a potential issue with mjfit's sim-fitting: if the peg exactly matches the circumscribed polygon of the hole, ground-truth behavior should be "jammed," yet mjfit lets it drop once yaw matches. He also notes the intuitive "not jammed" metric in videos requires larger clearance for other methods, and that mjfit's official benchmark always uses cylinder pegs for round holes—a very strong prior.
More from Embodied
- EgoXperience turns egocentric video into world-grounded 3D reconstructions in one forward pass — liuziwei7 · 2026-10-20
- Tesla ties key employee equity to Optimus production milestones in Oct 19 proxy filing — Scobleizer · 2026-10-20
- NVIDIA OmniDreams: real-time generative world model for closed-loop AV simulation, at 16 GB300 per video — yassineyousfi_ · 2026-10-20
- HandsOn Benchmark: 66-Task Dexterous Manipulation Test Shows Best Policy at 29% Success — Tianwei Luo · 2026-10-20
- Sparse Feature Unlearning Reduces State Hallucination in Vision-Language-Action Models — Jiho Lee · 2026-10-20
- LIBERO-Industry benchmarks robot manipulation under factory constraints — IAIE-PhysicalAI · 2026-10-20