NVIDIA's VeriFine Co-Evolves Policies and Judges for Self-Improving Embodied AI
drmapavone · x · 2026-10-20
NVIDIA researchers with UCLA, UC Berkeley and Stanford (including Marco Pavone) introduce VeriFine, an agentic harness for continuous self-improvement in Physical AI where policies, training curricula, and judges evolve together. As policies improve and develop new failure modes, static judges become a bottleneck—especially in embodied settings needing spatial grounding, causal reasoning, and safety awareness. VeriFine identifies policy failures, converts them into training signals, and when the judge itself bottlenecks progress, brings in targeted human feedback to refine its evaluations. Reported results: after three alternating judge/policy loops, policy reasoning score rises from 60.56 to 71.61 and judge-human Pearson correlation from 0.53 to 0.82 across autonomous driving and robot navigation tasks.
More from Embodied
- Berkeley EMBER humanoid seminar series: CMU's Guanya Shi on generalist humanoids — GuanyaShi · 2026-10-20
- Apple reportedly set to unveil touchscreen, eSIM MacBook on Oct 28 — xiaohu · 2026-10-20
- Same interface drives a real robot: buzz-wire demo and EmbodiedSafetyExam — ManlingLi_ · 2026-10-20
- RoboEngineer demo: AutoResearch on sim2sim — ManlingLi_ · 2026-10-20
- RoboDynamics demo: real-time tasks where the world won't pause — ManlingLi_ · 2026-10-20
- RoboARC demo: reasoning under spatial constraints — ManlingLi_ · 2026-10-20