AnyGoal: benchmarking world model generalization beyond expert rollouts
galilai-group · hf · 2026-10-20
Researchers from Brown, Mila, NYU and Cold Spring Harbor introduce AnyGoal, a benchmark testing whether world models can generalize dynamics to unseen combinations of familiar states and actions for planning. It spans 12 environments — TwoRoom navigation, Push-T variants, and robotic peg/cube manipulation — separating base training tasks from free-space and contact-rich inference tasks.
More from Embodied
- Cyborg routine: checking free memory on a JavaCard implant with a reader and Python — TechPreacher · 2026-10-20
- First Humanoid Robot Fashion Challenge brings designers and robotics researchers together — kscottz · 2026-10-20
- Robotics feud: mjfit generates convex decompositions for MuJoCo, no simulator needed — KyleMorgenstein · 2026-10-20
- HoliWorld: scalable auto-labeling pipeline turns egocentric video into 3D/4D data — liuziwei7 · 2026-10-20
- mjfit peg benchmark under fire: cylinder-peg setup embeds a strong prior — YouJiacheng · 2026-10-20
- "Not jammed" isn't the metric: mjfit author clarifies benchmark measures contact onset and play — YouJiacheng · 2026-10-20