Humanoid RL Shows Scaling Laws: Picking Throughput Jumps 85%
lukas_m_ziegler · x · 2026-08-04
Reinforcement Learning Shows Scaling Laws in Robotics
UK startup @TheHumanoidAI revealed KinetQ Ascend, running real-world reinforcement learning (RL) on production humanoid systems. The key finding is that robot performance scales predictably with training time, similar to LLM scaling laws, suggesting a path to 100% reliability.
- Bin-picking task: RL increased throughput by 42% at 1.5x human demo speed.
- Cluttered tote picking: After just a few days of training, throughput jumped 85% and success rates went from 80% to 98%.
Reimagine Robots Emerges from Stealth
Founded by Jonathan Scholz, who built Google DeepMind's Applied Robotics team in London, @reimaginerobots announced its emergence from stealth.
- Approach: Show the robot a task, watch it try, and correct it on the spot without needing specialist programmers.
- Deployment: Already used in manufacturing and electronics disassembly, reducing the time to prototype a new robot behavior from 1 day to 10 minutes.
More from Embodied
- Robot North to Work Full Shifts at Real DQ Store in Shanghai — ZeYanjie · 2026-08-04
- Peking University Proposes Data Pyramid Framework for Embodied Manipulation — 机器之心 · 2026-08-04
- Expressive Robots Trigger Hormonal Changes, But Errors Cause Harder Trust Collapse — CyberRobooo · 2026-08-04
- Humanoid Robots Don't Need Perfection to Disrupt the Economy, Just Cost Efficiency — VraserX · 2026-08-04
- Procore Acquires DroneDeploy for $845M to Build Construction AI — lukas_m_ziegler · 2026-08-04
- G_Bionics Unveils Gene.01 Humanoid with Full-Body Tactile Smart Skin — lukas_m_ziegler · 2026-08-04