Reproduce “AdaTracker: Learning Adaptive In-Context Policy for Cross-Embodiment Active Visual Tracking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kui Wu, Hao Chen, Jinzhu Han, Haijun Liu, Churan Wang, Yizhou Wang, Zhoujun Li, Si Liu and 1 more
AdaTracker is a framework that enables a single policy to perform adaptive active visual tracking across diverse robotic platforms in a zero-shot manner by implicitly inferring embodiment-specific constraints. The system achieved a mean success rate of 0.91 across unseen simulated environments and successfully adapted to real-world wheeled, quadruped, and drone robots without retraining.
Key insights
- Implicitly inferring and encoding embodiment-specific constraints from interaction history provides a more effective and adaptable conditioning signal for cross-embodiment policies than explicit descriptors.
- Auxiliary tasks, such as predicting immediate reward and camera height, significantly stabilize and enhance the relevance of the learned latent embodiment context for robust policy control.
- Conditioning a recurrent control policy on a dynamically inferred and stable embodiment context enables robust zero-shot generalization across heterogeneous robotic platforms with varying physical characteristics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.