RoboArchitect automates robotic agent harness with persistent memory
Original title: RoboArchitect: Automating the Design of Robotic Agents with Persistent Memory
Coding model iteratively evolves harness around frozen policies raising success rates 3-4x on unseen scenes and tasks.
RoboArchitect observes robot rollouts, diagnoses failures and iteratively revises the harness around a frozen policy. The harness evolves on a set of episodes for optimization and is evaluated on held-out scenes and tasks. Without updating any model weights the approach raises average success from 9.7% to 43.3% on unseen scenes and from 12.7% to 40.2% on unseen tasks. The evolved harness carries learned experience such as effective grasp depths into later decisions.
- Raises success 4.5x on unseen scenes
- Raises success 3.2x on unseen tasks
- Automates harness design without finetuning
- Includes persistent memory of effective strategies
- Largest gains where benchmarks differ from training data
From the original article
1 min read at huggingface.co
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. Yet today these harnesses are built by hand and redesigned for each new robot or benchmark. We show that harness design can be automated, offering a strong alternative to finetuning on a novel setup.
By Shengguang Wu, Suning Huang, Cherie Ho · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.