NTU's RoboFoundry evolves the robot software stack itself, lifting frozen GPT-5.5 from 56.9% to 72.7%
jiqizhixin · x · 2026-10-19
Nanyang Technological University presents RoboFoundry (arXiv 2609.32862), a system that lets the entire robot software stack autonomously evolve — no model parameter changes required. On EmbodiedBench, frozen GPT-5.5 jumps from 56.9% to 72.7%; all five base models improve (GPT-6 Astra +6.1, Qwen3.7-Plus +10.3, GLM5.3-Flash +7.4, Qwen3.8-27B +14.2).
Key ideas:
- "System-as-Policy": an embodied policy is the model plus memory, skills, interaction harness and code-as-policy logic; most work improves one component in isolation and freezes the rest after deployment.
- An "Act — Reflect — Repair — Promote" loop executes across embodiments, attributes failures to capabilities, edits the task-level system, evaluates candidates, and promotes changes that generalize.
- On long-horizon memory benchmark RoboMemArena it reaches 72.8 success rate, beating PrediMem, MemER, HiF-VLA and others; LIBERO-PRO robustness hits 92.0/91.8, with real-world adaptation shown on physical robots.
More from Embodied
- Chinese blogger builds working electric hoverboard with 10 motors and real-time AI voice coaching — churchkey · 2026-10-19
- Security robot dog enforces parking at Atlanta condo complex, sparks surveillance backlash — 4KTV · 2026-10-19
- The "Silly Name" heuristic: Kiva, Symbotic, Otto and other robotics exits all failed to profit — MatthewChang · 2026-10-19
- Robot learns to stay upright by shifting mass while keeping feet planted — first checkpoint — _Stocko_ · 2026-10-19
- Jensen Huang: physical AI is the next wave, factories to be highly robotic within 10 years — AryHHAry · 2026-10-19
- What 2+ Hours of Chinese Factory Footage Teaches About Humans in Automated Plants — HaozhiQ · 2026-10-19