Pretraining Makes Robot Policies Reliable but Hurts Posttraining, New Paper Finds
siddkaramcheti · x · 2026-11-05
A new paper by Zhanyi et al., in collaboration with Shuran Song, shows that continued supervised pretraining of generative robot policies improves reliability at the expense of coverage — a tradeoff that directly affects downstream posttraining. More pretraining can raise success rates while making policies harder to fine-tune.
The paper explains why this happens and how to preserve coverage. Max Simchowitz notes that LLM research has perplexity as a metric correlating with posttraining performance, but robotics still lacks an analogue — and calls this work a strong first step toward better pretrained-checkpoint evaluation metrics.
More from Embodied
- PhD student blogs on challenges of latent action models for robotics pretraining — EugeneVinitsky · 2026-11-05
- Voxel-based sim2real author flags two hurdles: voxel resolution vs speed and data scarcity — chris_j_paxton · 2026-11-05
- Liquidating ~18 humanoid robots in SF: Unitree G1/H1/R1, AGIbot, Booster, EngineAI in stock — cixliv · 2026-11-05
- Nothing CEO says mysterious AI-first device launches next month — TechCrunch AI · 2026-11-05
- XR Blocks team runs talk, demos and 'World as a Canvas' workshop at UIST 2026 — DuRuofei · 2026-11-05
- YC-backed Ishiki Labs launches Fern, a robotics sim platform that builds high-fidelity sims from 3 teleop episodes — ycombinator · 2026-11-05