SEAR: Self-Evolving Agentic Robotics
SEAR: Self-Evolving Agentic Robotics
The paper presents SEAR, a framework and benchmark for robotic agents that autonomously improve through self-evolving strategies. It evaluates how language models can write control code, test it, and retain skills for future tasks. The study explores learning dynamics, transferability of control code across tasks and simulators, and model-specific strategies for task execution, using seven language models across 330 simulated and five physical tasks over four years of runtime.
This paper introduces SEAR, a framework and benchmark for self-evolving agentic robotics, which explores how language models can autonomously improve robotic performance through code writing, testing, and skill retention. The study highlights key findings on learning dynamics, transferability of control code, and model-specific strategies for task execution.
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.
Reported numbers
simulated tasks
330
View original evidence
budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second,
Open source S4physical tasks
5
View original evidence
budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second,
Open source S4agent runtime
>4
View original evidence
budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second,
Open source S4frontier language models
7
View original evidence
budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second,
Open source S4simulator
1
View original evidence
control code and shared routines transfer across tasks and simulators, while task-specific solutions do not. On a real robot, we observe a performance gap between agent-designed practice and novel human-designed scenes. Third, different models evolve more effectively in different ways. Some do best acting live, others writing code first. And a text-only model keeps pace with a comparable multimodal one: perception he
Open source S5real robot
1
View original evidence
control code and shared routines transfer across tasks and simulators, while task-specific solutions do not. On a real robot, we observe a performance gap between agent-designed practice and novel human-designed scenes. Third, different models evolve more effectively in different ways. Some do best acting live, others writing code first. And a text-only model keeps pace with a comparable multimodal one: perception he
Open source S5
- code: not_reported
- dataset: not_reported
- paper: not_reported
- weights: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second,
Open source S4
ks, in a new simulator, and on a real robot? The framework runs an agent in three modes: it writes code that becomes the policy (code-as-policy), acts live in the control loop (agent-as-policy), or works as a researcher that fine-tunes a vision-language-action model (auto-research). The benchmark adds controls that separate retained experience from repeated attempts, familiar tasks, privileged simulator state, larger
Open source S3
control code and shared routines transfer across tasks and simulators, while task-specific solutions do not. On a real robot, we observe a performance gap between agent-designed practice and novel human-designed scenes. Third, different models evolve more effectively in different ways. Some do best acting live, others writing code first. And a text-only model keeps pace with a comparable multimodal one: perception he
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org