EqC paper: introspective robot policies that know when to ask for help
du_yilun · x · 2026-11-12
A new arXiv paper (2610.18790) introduces Equilibrium Energy Control (EqC), a framework for introspective lifelong robot learning built on an explicit conditional energy landscape.
- One energy landscape drives both action generation and introspection: the geometry of action refinement reveals whether behavior stays consistent with the learned landscape, so the policy itself knows when to ask for help.
- Energy-Gated DAgger × Astra creates a self-improvement loop: detect a competence gap → agent recovery → learn from recovery trajectories → expand autonomy, claimed more effective than human-judged interventions.
- Integrates with Diffusion Transformer and flow-matching VLA policies; matches or improves performance in simulation and real bimanual manipulation.
- Real-world garment folding went from 0% to 100% autonomous success after just 6 agent-generated interventions, a practical path toward self-improving robots.
More from Embodied
- Scaling H2R workshop at CoRL 2026 probes scaling laws for human-to-robot learning — mangahomanga · 2026-11-12
- Stanford's FIDO robot dog learns physics by playing fetch with humans — kevin_zakka · 2026-11-12
- iPhone Duo fold test hits record 413,000 cycles in 13 days with zero failures — SumitGup · 2026-11-12
- kevin_zakka adds tactile sensing support to mjlab robotics framework — kevin_zakka · 2026-11-12
- MCP Display turns an old Android phone into a programmable screen for AI agents — PatrickCavalcanti · 2026-11-12
- Dennis Hong showcases THOR hazard-response robot and ARTEMIS at AIVEX Frontier 2026 — DennisHongRobot · 2026-11-12