SEU's LeaP: a learnable source prior boosts robot manipulation success by 25.5 points
机器之心 · wechat · 2026-10-03
Wei Xiucai's team at Southeast University proposes LeaP (Learnable source Prior), accepted at CoRL 2026. Instead of starting generative robot policies (diffusion/flow matching) from a standard Gaussian, a lightweight proprioception-driven MLP head (0.21M params) jointly predicts the mean and state-adaptive variance of a diagonal Gaussian source distribution for action generation. On RoboTwin's 15 bimanual tasks it reaches 81.6% average success, +25.5 points over the standard-Gaussian baseline, and 80.0% on three real Franka tasks. Ablations show proprioception-only priors beat visual variants; learning mean plus adaptive variance (85.3%) clearly beats mean-only (78.0%) or fixed noise (62.7%); explicit likelihood and contrastive supervision add complementary gains. The prior transfers across generators (flow matching, diffusion bridge), and code is open-sourced.
More from Embodied
- Berkeley's Humanoid Intelligence Center wins both tracks at IROS 2026 RoCo Challenge — berkeley_ai · 2026-10-03
- Neuralink pretrains on 50,000 hours of neural data, cuts calibration to 10 minutes — XFreeze · 2026-10-03
- CoRL 2026 to host Sim-to-Real-to-Field workshop on transfer science, CFP open — berkeley_ai · 2026-10-03
- Yann LeCun: two years later we're still far from human-level AI — your cat beats it at physics — ylecun · 2026-10-03
- VRChat fluid sim adds collider buoyancy, performance presets and Japanese localization — Michael_Moroz_ · 2026-10-03
- Tesla Robotaxi in Austin extends service hours to 11pm — yunta_tsai · 2026-10-03