Chelsea Finn says robotics is bottlenecked by rollout cost and memory
chris_j_paxton · x · 2026-07-28
Chelsea Finn: robot scaling is bottlenecked by rollout cost and missing memory
The post highlights two key constraints in robotics foundation models:
- RL scaling is expensive in the physical world. A million 1-minute trajectories would take roughly 700 robot-days, so scaling PPO/GRPO-style post-training is fundamentally different from LLM training.
- Memory is still missing. Most SOTA robot systems have effectively no memory, and naively extending video history is too costly: 10 seconds at 50 Hz across 4 cameras is about 500K tokens.
The cited approach uses short-term video memory plus long-term text summaries of the prior 10–15 minutes, which helped enable a fully autonomous multi-step kitchen workflow.
Related event: Chelsea Finn Identifies Physical Rollout Costs as RL Bottleneck(2 posts)→
More from Embodied
- GEFIT uses a RealSense D555 stereo camera for robotic pick-and-place — chrismatthieu · 2026-07-28
- Does an AI agent really need dedicated hardware, or just better software? — Jayden_Wong0211 · 2026-07-28
- SolidWorks build turns seven actuators into a 7-DOF humanoid robot arm — philfung · 2026-07-28
- Expanding Flow Maps can insert tokens or nodes while denoising — KevinKaichuang · 2026-07-28
- A cracked pair of Ray-Ban glasses becomes a joke about smart-glasses backlash — AIFlow_ML · 2026-07-28
- Physical AI could make industrial robots easier to deploy — Responsible-Grass452 · 2026-07-28