RoboArchitect: Automating the Design of Robotic Agents with Persistent Memory
arXiv cs.AI · 31 Oct 2026, 03:58 AEST
RoboArchitect automates the design of robotic agents with persistent memory, significantly improving their performance. The system enhances success rates on unseen scenes and tasks without updating model weights.
Key facts
- Recent work shows that a frozen robot policy performs far better inside an agent harness, where control code decides when to query the policy, consult vision tools, or command the arm directly.
- In RoboArchitect, a coding model observes a robot's rollouts, diagnoses failures, and iteratively revises the harness around the frozen policy.
- Without updating any model weights, RoboArchitect raises average success from 9.7% to 43.3% on unseen scenes (4.5x) and from 12.7% to 40.2% on unseen tasks (3.2x), across multiple benchmarks and policies.
- The harness also evolves its own persistent memory, carrying learned experience, such as effective grasp depths, into later decisions.
Read aloud
RoboArchitect automates the design of robotic agents with persistent memory, significantly improving their performance. The system enhances success rates on unseen scenes and tasks without updating model weights. Recent work shows that a frozen robot policy performs far better inside an agent harness, where control code decides when to query the policy, consult vision tools, or command the arm directly. In RoboArchitect, a coding model observes a robot's rollouts, diagnoses failures, and iteratively revises the harness around the frozen policy. Without updating any model weights, RoboArchitect raises average success from nine point seven percent to forty three point three percent on unseen scenes four.5x and from twelve point seven percent to forty point two percent on unseen tasks three.2x, across multiple benchmarks and policies. The harness also evolves its own persistent memory, carrying learned experience, such as effective grasp depths, into later decisions.
Sources in this cluster
- RoboArchitect: Automating the Design of Robotic Agents with Persistent Memory
arXiv cs.AI · 31 Oct 2026, 03:58 AEST