EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution
EmbodiedRSI is pitched as a robot-learning harness that spends physical trials on experiments meant to separate competing code and skill hypotheses.
The paper says the system keeps those hypotheses in a graph, chooses robot interactions by value of information, then uses the outcomes to update both code and skills. It reports 77.0% overall success on RoboCasa365, including 71.3% on Composite-Unseen, versus 40.1% for the best baseline. It also reports 86.8% overall success on LIBERO-Pro and 71.3% zero-shot success on a real-world robot across several tasks. ArXiv · AI/CL/LG's note
The paper says the system keeps those hypotheses in a graph, chooses robot interactions by value of information, then uses the outcomes to update both code and skills. It reports 77.0% overall success on RoboCasa365, including 71.3% on Composite-Unseen, versus 40.1% for the best baseline. It also reports 86.8% overall success on LIBERO-Pro and 71.3% zero-shot success on a real-world robot across several tasks. ArXiv · AI/CL/LG's note
score 5