Learning and Transferring Closed-Loop Robot Software
Learning and Transferring Closed-Loop Robot Software
This research investigates whether improvements in closed-loop robot policy implementations for source tasks can be transferred to new tasks. By treating complete closed-loop implementations as reusable execution experience, the authors use a coding agent to generate and iteratively improve policy code from demonstrations and simulation feedback. Validated implementations are stored in a software archive. The study evaluates policy learning across four RoboCasa source tasks and cross-task transfer on nine additional target tasks. Results show that iterative optimization increases success rates from 28.3% to 64.2% on source tasks, and using archived code improves performance on new tasks from 41.5% (initially generated) to 57.0%. The approach is further validated on a physical Unitree G1 robot.
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
Reported success rate
28.3%
View original evidence
es using archived implementations, target demonstrations, and execution feedback. We evaluate policy learning and its relationship to cross-task transfer. For learning, across four source tasks in RoboCasa, iterative optimization increases the success rate from 28.3% with initially generated policy code to 64.2% with validation-selected policy code. We further validate this learning approach on a physical Unitree G1.
Open source S4Reported success rate
64.2%
View original evidence
es using archived implementations, target demonstrations, and execution feedback. We evaluate policy learning and its relationship to cross-task transfer. For learning, across four source tasks in RoboCasa, iterative optimization increases the success rate from 28.3% with initially generated policy code to 64.2% with validation-selected policy code. We further validate this learning approach on a physical Unitree G1.
Open source S4Reported success rate
45.2%
View original evidence
For transfer, across nine additional target tasks in RoboCasa, success rates are 45.2% without references, 41.5% using initially generated policy code from source tasks, and 57.0% using their validation-selected policy code. These results show that improving source-task code also improves learning on new tasks.
Open source S5
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
es using archived implementations, target demonstrations, and execution feedback. We evaluate policy learning and its relationship to cross-task transfer. For learning, across four source tasks in RoboCasa, iterative optimization increases the success rate from 28.3% with initially generated policy code to 64.2% with validation-selected policy code. We further validate this learning approach on a physical Unitree G1.
Open source S4
For transfer, across nine additional target tasks in RoboCasa, success rates are 45.2% without references, 41.5% using initially generated policy code from source tasks, and 57.0% using their validation-selected policy code. These results show that improving source-task code also improves learning on new tasks.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org