Chelsea Finn says robot RL is bottlenecked by physical rollout cost, not algorithms
ycombinator · x · 2026-07-27
Chelsea Finn’s YC Startup School talk argued that robot RL is constrained more by physical rollout cost than by algorithm quality. She estimated that 1 million 1-minute trajectories would take roughly 700 robot-days, making LLM-style scaling dynamics fundamentally different for robotics.
Key points from the talk:
- Memory is a major missing piece in robot foundation models. Most state-of-the-art systems still have effectively no memory, while naive video history is too expensive.
- π’s approach uses short-term video memory plus long-term text summaries of the previous 10–15 minutes, which helped it complete a fully autonomous multi-step kitchen-cleaning task.
- Finn also suggested generalist robot models may already be outperforming specialist pipelines, by mixing demos, rollouts, human video and web data, then using richer prompting and metadata.
Related event: Chelsea Finn Identifies Physical Rollout Costs as RL Bottleneck(2 posts)→
More from Embodied
- XPENG's IRON robot demos full-duplex speech with 9-mic array and lip reading — ChrisGPT · 2026-09-23
- Uber riders can now hail Waymo robotaxis on Austin freeways — ATTlKA · 2026-09-23
- M5 Ultra LLM test: 4x faster prompt processing, but double the power draw — DigitalguyCH · 2026-09-23
- Figure CEO: Humanoid Robotics Needs Four Stages and Eventually Hundreds of Billions — adcock_brett · 2026-09-23
- InstinctFlash: open-source runtime runs 8 robot model families in real time on one commercial GPU — chris_j_paxton · 2026-09-23
- M5 Ultra 96GB vs RTX 5090-6000 Pro: which rig for local video generation at $6.5K-$27K — MaxwellHusk · 2026-09-23