Robot imitation learning plateaus with more data, pushing the field toward RL
chris_j_paxton · x · 2026-11-12
Chris Paxton shares a CoRL talk slide showing imitation learning plateaus as demonstration data grows: shirt-folding throughput went 1x (May 2024) → 1.5x (Aug) → 1.1x (Oct), bussing 1x → 4x → 1.3x. The marginal benefit of more demonstrations stops improving performance — the core reason the speaker (Kay) concluded the field needs reinforcement learning / target finetuning to keep pushing capability.
More from Embodied
- Ex-Tesla Autopilot simulation leads exit stealth at $1.1B valuation with $160M raised — heyshrutimishra · 2026-11-12
- MIMIC and BMW train video-action model hitting up to 100% success on industrial tasks — pess_r · 2026-11-12
- Simular launches box, a dedicated computer built for running AI agents — _akhaliq · 2026-11-12
- Dex hand hits 99% on rubber-band task — hardware, not AI, is the bottleneck — chris_j_paxton · 2026-11-12
- Dyna co-founder Jason Ma keynotes on scaling real-world robot foundation models — chris_j_paxton · 2026-11-12
- Roboticist: RL is mandatory for robots too, behavior cloning was a detour — chris_j_paxton · 2026-11-12