Recoleta
GitHub
Source note

NavWAM: A Navigation World Action Model for Goal-Conditioned Visual Navigation

Authors
Daichi Azuma, Taiki Miyanishi, Koya Sakamoto, Shuhei Kurita, Yaonan Zhu, Petr Khrapchenkov, et al.
Source
arXiv
Published
Collection
Embodied AI
Visual NavigationWorld ModelDiffusion PolicyGoal Conditioned ControlRobot Learning
Open arXivMarkdown source

Summary

NavWAM solves goal-conditioned visual navigation by turning future prediction into direct action, so the robot can use visual foresight without a separate planner. It matters because navigation under partial observability needs both next-view prediction and action selection, and prior world models split those jobs.

Problem

  • Goal-conditioned image navigation must work from partial egocentric views, so the robot needs to predict how motion changes what it sees and whether that motion moves it toward the goal.
  • Prior navigation world models predict future views, but they still depend on CEM-style planning or another external selector to choose actions.
  • Direct navigation policies output actions efficiently, but they do not explicitly learn future egocentric prediction and goal-progress estimation in one model.

Approach

  • NavWAM is a diffusion-transformer policy built on Cosmos Predict2 that treats navigation as denoising one latent canvas.
  • The canvas contains observed frames plus prediction frames for an action chunk, a future state, future egocentric views, and a goal-progress value.
  • The model learns future view prediction, action generation, and value estimation together, instead of using separate heads or an external planner.
  • At test time, it runs in policy mode and outputs an action chunk directly in a receding-horizon loop.
  • It also supports optional best-of-N sampling, but the main results use default policy mode without CEM-style search.

Results

  • On go stanford, NavWAM beats planning baselines in trajectory error: ATE 0.324 vs 0.453 for NWM and 0.455 for Cosmos Predict2; RPE 0.099 vs 0.107 and 0.109.
  • With fine-tuning on go stanford, NavWAM reaches ATE 0.192 and RPE 0.070, which the paper reports as its best navigation result on that benchmark.
  • Future prediction stays useful after turning the model into a policy: subject consistency rises from 0.524 for NWM to 0.668 for NavWAM before fine-tuning, and stays at 0.635 after fine-tuning.
  • On sit, NavWAM is competitive with OmniVLA despite using a 2B video backbone rather than OmniVLA's 7B OpenVLA backbone; it gets ATE 0.077 vs 0.086 at h=4 and 0.144 vs 0.162 at h=8, with SR 46.3% vs 45.4% at h=4 and 15.9% vs 12.1% at h=8.
  • In real-robot tests on 24 episodes, NavWAM reaches the goal in 19/24 episodes, or 79.2%, compared with 14/24 for OmniVLA and 4/24 for NWM.

Link