NavThinker: Action-Conditioned World Models for Coupled Prediction and Planning in Social Navigation

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• NavThinker achieves state-of-the-art navigation success rates in the Social-HM3D environment.result 90%
• NavThinker demonstrates zero-shot transfer capability from Social-HM3D to Social-MP3D environments.result 85%
• NavThinker successfully deployed on a real-world Unitree Go2 robot, validating practical applicability.result 90%
• NavThinker uses an action-conditioned world model operating in Depth Anything V2 patch feature space.technical 95%
• NavThinker trains policies with DD-PPO while injecting world-model think-ahead signals through two mechanisms.method 90%
• NavThinker's performance relies on accurate sensory input and environmental understanding, which can be affected by real-world conditions.limitation 85%
• NavThinker addresses the coupled prediction-planning challenge where robot actions and human motion mutually influence each other.technical 90%
• NavThinker's world model produces future depth maps and human trajectories through multi-head decoders.method 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card