LeaP from Southeast University, accepted at CoRL 2026, makes diffusion action priors state-aware
jiqizhixin · x · 2026-10-11
- Southeast University's Wei Xiushen team presents LeaP (Learnable source Prior), accepted at CoRL 2026, tackling how diffusion-based action generation should model its starting point.
- Standard diffusion action generation starts from an observation-agnostic Gaussian; methods like A2A and VITA use proprioceptive or visual cues but treat the starting point as deterministic without modeling uncertainty.
- LeaP replaces the fixed standard Gaussian with a proprioceptively conditioned diagonal Gaussian, jointly learning mean and variance so the generation start adapts to the robot's state.
- On 15 RoboTwin simulated manipulation tasks, LeaP reaches an average success rate of 81.6% (full numbers truncated in the post).
More from Embodied
- AI Passport turned pocket EDC: 27 games, 16 tools, offline-first drift-bottle social — ezshine · 2026-10-11
- DeformX co-simulation framework for deformable objects lands IROS 2026 oral — rsasaki0109 · 2026-10-11
- Eric Jang resurfaces 2021 essay: bet on compute pressure sparking spontaneous intelligence — ericjang11 · 2026-10-11
- Wuji opensources mjlab PPO stack for dexterous hand: in-hand reorientation and pen spinning — rohanpaul_ai · 2026-10-11
- First Waymo ride to SFO: 10/10 experience, never hailing Uber again — braelyn_ai · 2026-10-11
- Lifesize R2D1-driven skeleton animatronic roasts trick-or-treaters on Halloween — natesiggard · 2026-10-11