Top-labs survey: 5 of 7 robot self-improvement studies show no cross-task gains, zero confirmed L5
rohanpaul_ai · x · 2026-10-15
A survey on Physical Recursive Self-Improvement from MIT, Stanford, Google, Microsoft and other top labs finds that robot self-improvement still leans heavily on humans for data collection, failure diagnosis, and update selection.
Key findings:
- 5 of the 7 robot self-improvement studies reviewed never demonstrated gains on a different task, and most evaluation protocols lack cross-task tests.
- The paper defines a 6-level framework (L0–L5), separating capability-level from meta-level updates and within-task from across-task gains, reserving L5 for independently verified recursive gain.
- A map of systems from 2022–2026 contains no confirmed L5 case.
- Real physical trials are slow, noisy, costly, sometimes irreversible, so digital self-improvement methods need rethinking. The survey concludes current literature does not yet demonstrate Physical RSI as defined.
More from Embodied
- Lukas Ziegler joins SPRIND Next Frontier Robotics Challenge jury, backs modular thesis for Europe — lukas_m_ziegler · 2026-10-15
- Gothic Church 3D Gaussian Splat Coming to Vision Pro, Hailed as Sharpest Ever — willeastcott · 2026-10-15
- Robotics IPOs trade like memecoins: +12% day-one median in China, now 13% below offer — LexSokolin · 2026-10-15
- Etherbulb turns a $15 ESP32-C3 smart bulb into an SDR that paints the room's radio — IgorCarron · 2026-10-15
- Simulate before you begin: validate walking policies with a toy model before CAD — IgorCarron · 2026-10-15
- Gemini controls a real robot from one prompt: SSH in, find joints, learn from demo — heiga_zen · 2026-10-15