AdaRSI
An on-demand recursive self-improvement framework for embodied agents. It attributes the source of failures to decide whether to optimize the tool rack or fine-tune the actuator, achieving precise and efficient capability improvement.
Timeline (last 90 days)
AdaRSI proposes an on-demand recursive self-improvement framework that treats "when to change the weights" as an explicit decision problem.
The AdaRSI pipeline consists of three steps: first improve the coding agent's skills and tool harness, attribute persistent failures, and fine-tune only the sub-steps that are genuinely unreliable.
In the attribution step, AdaRSI retries from saved states to determine whether the failure stems from how it was invoked or from execution capability.
AdaRSI's fine-tuning demonstration data is collected by the system itself and reuses experience accumulated from previous tasks.
The paper validated AdaRSI on all 50 target tasks of RoboCasa365, using GR00T N1.5 as the executor and without using any human demonstrations throughout.
On unseen composite tasks, AdaRSI increased the success rate from 4.6% to 13.8%, roughly a threefold relative improvement.
Compared with the approach of "fine-tuning every sub-step", AdaRSI achieved a slightly higher success rate while consuming only 28% of the GPU time and 22% of the self-collected demonstration data.
The approach with fine-tuning but without tool harness optimization, despite a larger training volume, achieved far less than AdaRSI.
AdaRSI currently still relies on the mechanism of retrying from saved states, which holds in environments where states are reproducible, but in highly dynamic or irreversible real-world scenarios, the attribution cost may rise significantly.
All Facts (9)
AdaRSI proposes an on-demand recursive self-improvement framework that treats "when to change the weights" as an explicit decision problem.
50%UnverifiedThe AdaRSI pipeline consists of three steps: first improve the coding agent's skills and tool harness, attribute persistent failures, and fine-tune only the sub-steps that are genuinely unreliable.
50%UnverifiedIn the attribution step, AdaRSI retries from saved states to determine whether the failure stems from how it was invoked or from execution capability.
50%UnverifiedAdaRSI's fine-tuning demonstration data is collected by the system itself and reuses experience accumulated from previous tasks.
50%UnverifiedThe paper validated AdaRSI on all 50 target tasks of RoboCasa365, using GR00T N1.5 as the executor and without using any human demonstrations throughout.
50%UnverifiedOn unseen composite tasks, AdaRSI increased the success rate from 4.6% to 13.8%, roughly a threefold relative improvement.
50%UnverifiedCompared with the approach of "fine-tuning every sub-step", AdaRSI achieved a slightly higher success rate while consuming only 28% of the GPU time and 22% of the self-collected demonstration data.
50%UnverifiedThe approach with fine-tuning but without tool harness optimization, despite a larger training volume, achieved far less than AdaRSI.
50%UnverifiedAdaRSI currently still relies on the mechanism of retrying from saved states, which holds in environments where states are reproducible, but in highly dynamic or irreversible real-world scenarios, the attribution cost may rise significantly.
50%