Few-Shot Residual Adaptation of Pretrained Latent World Models
Few-Shot Residual Adaptation of Pretrained Latent World Models
This paper introduces a few-shot residual adaptation approach for pretrained latent world models that learns self-supervised corrections from latent prediction errors without modifying pretrained weights. The method applies corrections inside the latent predictor and directly to its predicted latent state while keeping the model frozen. Evaluated in simulation and on a real robot under center-of-mass and friction shifts, the approach recovers 90% success under center-of-mass shift and 100% success under modified friction using three interaction episodes for adaptation.
The authors evaluate a method for adapting frozen pretrained latent world models to physical dynamics changes using few-shot residual adaptation, reporting recovery of task success rates on a real robot under specific friction and center-of-mass shift conditions.
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.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
self-supervised corrections from latent prediction errors. The corrections are applied both inside the latent predictor and directly to its predicted latent state, while the pretrained world model remains frozen. We evaluate the approach in simulation and on a real robot under center-of-mass and friction shifts. These changes reduce the success of the frozen model to near zero. Using three interaction episodes for ad
Open source S3
aptation, our dual-site approach recovers 90\% success under the center-of-mass shift and 100\% under the modified friction on the real robot. Ablations show that correcting both locations outperforms single-site adaptation. Our diagnostics further suggest that much of the learned correction reduces a general prediction mismatch, while successful control also depends on shift-specific information captured during adap
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org