Hydra: A Navigation World Action Model with Discrete Latent Planning and Continuous Flow-Matching Execution
Abstract Overview
Hydra is a navigation world action model designed to address the mismatch between generative world models and external planners that makes real-time control difficult. It builds a unified latent manifold over visual observations, robot poses, and actions, then compresses these modalities with vector-quantized bottlenecks so planning can be performed directly in a discrete latent space. The method uses Discrete Latent Planning to search over learned kinodynamic intents and rank candidates with a kinematic-perceptual cost, avoiding per-candidate pixel decoding. For execution, Hydra maps the selected discrete intent to smooth continuous trajectories through flow matching, and the paper evaluates the system on physical robots for planning and path-following tasks.
Novelty
The paper’s main distinction is combining discrete latent planning with continuous flow-matching execution inside a single navigation world action model. Its planner searches and evaluates candidate trajectories within a shared latent manifold using vector-quantized intent and visual codebooks, rather than relying on continuous external sampling and image-space scoring.
Results
In video prediction, Hydra achieves better temporal-coherence FVD than the reported NWM baseline while remaining much faster than NWM for generating 100 clips. In physical close-range planning, Hydra runs in about 0.9 seconds per plan versus over 500 seconds for NWM/B, and it attains 10/10 success in unobstructed trials and 8/10 in both occluded and blind-corner scenarios. In local path following on Spot, Hydra’s guidance variants match or exceed reactive baselines in several scenarios despite using sparser waypoint guidance, though its full planning mode is weaker on the long-distance scenario.
Key Points
- Hydra unifies vision, pose, and action into a shared latent representation and discretizes future intents, enabling planning directly over learned kinodynamic primitives.
- Its kinematic-perceptual cost uses latent-space geometric, prior, entropy, and quantization signals to score safety and progress without decoding every candidate to pixels.
- Experiments on physical robots show substantial planning-speed gains over continuous world-model baselines and competitive path-following performance relative to reactive navigation policies.