Reproduce “Trajectory Learning with Graph Representations for Social Robot Navigation” 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
Berke Kartal, Burcu Kilic, Yigit Yildirim, Emre Ugur
A framework for social robot navigation combines a Graph Feature Autoencoder (GFAE) to represent complex crowd interactions and an extended Conditional Neural Movement Primitives (CNMP) module for trajectory generation. This approach mitigates error accumulation and implicitly incorporates anticipated pedestrian movements, leading to fewer intrusions into human personal zones and significantly reduced collisions in both simulated and real-world environments.
Key insights
- Decoupling crowd representation learning with a GFAE allows for the creation of robust, transferable, fixed-size embeddings that capture complex social interactions regardless of crowd density.
- Employing a trajectory-level learning objective in the CNMP module effectively mitigates the error accumulation typically seen in single-step imitation learning methods.
- Jointly predicting future crowd states and robot trajectories from a shared latent space enables implicit anticipation of pedestrian movements, facilitating preemptive social navigation without explicit, cascaded prediction networks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.