EPO at CoRL 2026: energy-based orchestration cuts humanoid policy switches 7x
du_yilun · x · 2026-11-08
Yilun Du's group presents Energy-Based Policy Orchestration (EPO) at CoRL 2026, a framework for orchestrating humanoid whole-body skills via per-policy energy values (with Pieter Abbeel, Leslie Pack Kaelbling, Tomás Lozano-Pérez et al.).
- Planners run on task timescales while feasibility changes on control timescales; EPO bridges this with a high-frequency mid-level layer: each low-level policy gets a differentiable, comparable energy measuring goal feasibility and expected attainment, optimized at control rate.
- Unlike a learned MoE router, EPO is compositional — adding a skill means training one new energy.
- Results: standing-manipulation tracking error drops 0.461 → 0.420 vs MoE-M4; policy switches fall 7x (2.08 vs 15.46 per episode); EPO-M4 survives 97%/93% of 3.0 m/s pushes that topple individual policies, with a 40 ms median time to first switch at 50 Hz control.
More from Embodied
- DeepMind's MuJoCo Physics Simulator Passes 16k GitHub Stars, Widely Used in Robotics — google-deepmind · 2026-11-08
- Agent Band's wrist-worn AI agent promo draws déjà vu mockery: 'wait, I've seen this one' — damnGruz · 2026-11-08
- ORA: open-weight DeepSeek model zero-shots real robot arms at $0.35 per task — JunGao33210520 · 2026-11-08
- Dimensional-Funded ESP32 Treasure Hunt Workshop Pushes American Open Source in Robotics — StewartalsopIII · 2026-11-08
- Samsung, SK Hynix and Micron in talks to standardize fab robots for fully automated fabs by 2030 — zephyr_z9 · 2026-11-08
- The brain as the next UI: upcoming article charts BCI-meets-agentic-AI roadmap via earbuds — melnykowycz · 2026-11-08