A Controlled Study of Models, Skills, Feedback, and Curriculum in Agentic Robotics
A Controlled Study of Models, Skills, Feedback, and Curriculum in Agentic Robotics
This research paper presents a controlled simulation study on agentic robotics, investigating how coding agents can iteratively improve robot policies through simulated and real-world interactions. The study spans nine manipulation tasks, 631 experiment configurations, and 28K coding agent rollouts, systematically varying model capability, skill priors, feedback structure, and curriculum. Results indicate that the primary bottleneck is often not information access but interpreting it into correct policy updates. Success depends on agents' ability to attribute failures, reason about task-relevant spatial and geometric structures, and selectively restructure code to improve policies.
This study systematically evaluates the self-learning loop in agentic robotics by varying model capabilities, skill priors, feedback structures, and curriculum across nine manipulation tasks, identifying that effective failure attribution and spatial reasoning are critical bottlenecks for policy improvement.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S2
Reported numbers
experiment configurations
631
View original evidence
arXiv:2611.09030v1 Announce Type: new Abstract: Agentic robotics asks coding agents to implement and iteratively improve robot policies from simulated and real-world physical interactions. We study the self-learning loop of agentic robotics through a controlled simulation study spanning nine manipulation tasks, 631 different experiment configurations, and 28K coding agent rollouts, systematically varying model capabi
Open source S2coding agent rollouts
28,000
View original evidence
arXiv:2611.09030v1 Announce Type: new Abstract: Agentic robotics asks coding agents to implement and iteratively improve robot policies from simulated and real-world physical interactions. We study the self-learning loop of agentic robotics through a controlled simulation study spanning nine manipulation tasks, 631 different experiment configurations, and 28K coding agent rollouts, systematically varying model capabi
Open source S2manipulation tasks
9
View original evidence
arXiv:2611.09030v1 Announce Type: new Abstract: Agentic robotics asks coding agents to implement and iteratively improve robot policies from simulated and real-world physical interactions. We study the self-learning loop of agentic robotics through a controlled simulation study spanning nine manipulation tasks, 631 different experiment configurations, and 28K coding agent rollouts, systematically varying model capabi
Open source S2
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2611.09030v1 Announce Type: new Abstract: Agentic robotics asks coding agents to implement and iteratively improve robot policies from simulated and real-world physical interactions. We study the self-learning loop of agentic robotics through a controlled simulation study spanning nine manipulation tasks, 631 different experiment configurations, and 28K coding agent rollouts, systematically varying model capabi
Open source S2
A Controlled Study of Models, Skills, Feedback, and Curriculum in Agentic Robotics
Open source S1
lity, skill priors, feedback structure, and curriculum. The experiment data suggests that the bottleneck is often not the access to sufficient information, but interpreting it into the right policy update. Across these factors, a successful self-learning loop depends on how effectively agents attribute failures, reason task-relevant spatial and geometric structure, and selectively restructure code to improve policies
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org