Galbot proposes observational scaling law for physical recursive self-improvement in VLM agents
Galbot · hf · 2026-10-20
Galbot's new paper systematically studies the mechanisms, scaling behavior, and generalization of physical recursive self-improvement (RSI) in vision-language agents.
- Introduces PhysRSI-eval with five axes: First-Attempt Performance, Efficiency, Success, Budget, and Improvement-from-failure
- Evaluates multiple VLMs on 18 manipulation tasks under a fixed interaction budget
- Results motivate the Physical Barrel Hypothesis: different capabilities constrain different parts of improvement
- Observational scaling analysis predicts agent performance from public benchmarks and forecasts later progress from early physical feedback
- Workspace interventions show removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task reuse cases succeed earlier
More from Embodied
- Tsinghua and AgiBot release RoboValue, a sim-and-real benchmark for robotic value models — Shengbang Liu · 2026-10-20
- AI-assisted robotic glove helps people regain hand control and independence — CurieuxExplorer · 2026-10-20
- Harness Pro: a $299 desk screen to monitor AI agents across all your machines — dee_hw · 2026-10-20
- A fruit fly's 166k-neuron connectome is now driving real robots, no AI training needed — CurieuxExplorer · 2026-10-20
- Sequoia Podcast: Inside the Robot Data Factory Making Millions of Hours of Human Training Data — Training Data (Sequoia) · 2026-10-20
- iPhone Duo orders slip to January 2027; analyst predicts balance by December — mitchdeg · 2026-10-20