Online Adaptation to Latent Physical Properties for Learning from Demonstration
Online Adaptation to Latent Physical Properties for Learning from Demonstration
This paper presents a learning-from-demonstration approach that adapts manipulation actions by learning latent representations of physical properties. The method jointly trains a dynamics model and object-specific latent variables from demonstrations, using these to condition an imitation policy. During execution, the model and policy weights are fixed while latent representations are updated based on discrepancies between predicted and actual observations. Experiments with objects having different centers of mass show improved task success rates through online adaptation, but simulation results indicate limited generalization to unseen physical properties, with success rates around 30% for center-of-mass locations not present in the training demonstrations.
The authors propose a method that jointly learns dynamics models and object-specific latent representations from demonstrations to enable online adaptation of manipulation actions based on physical properties, though simulation results indicate limited generalization to unseen center-of-mass locations with success rates around 30%.
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.
sentations of physical properties can improve task success rates. However, in simulation, success rates on center-of-mass locations absent from the demonstrations remained around 30%, indicating limited adaptation to unseen physical properties. Project page: https://isri-aist.github.io/latent-dynamics-project-page/
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org