World models let robots imagine possible futures, but exploiting this capability for real-time planning is bottlenecked by a representation misalignment: generative models and planners operate on decoupled manifolds, requiring computationally expensive decoding of every candidate back to the high-dimensional observation space for evaluation. In this paper, we present HYDRA, a discrete World Action Model that tackles this by establishing a unified latent manifold over visual states, physical poses, and control actions. By compressing this manifold through modality-specific Vector-Quantized bottlenecks, HYDRA yields discrete vocabularies of kinodynamic intents and visual states. This enables Discrete Latent Planning (DLP), where candidates are sampled directly from the shared manifold and ranked by a Kinematic-Perceptual Cost within the discrete latent space. To bridge discrete planning with the continuous commands required for physical actuation, HYDRA pairs DLP with conditional Flow Matching to map selected intents to smooth execution trajectories.
Below is the planning performance evaluated in clear, occluded, and narrow corridor turn environments.
| Model | Search Space | Plan Time ↓ | Clear (SR%) ↑ | Occluded (SR%) ↑ | Corner Turn (SR%) ↑ |
|---|---|---|---|---|---|
| NWM | Continuous (CEM) | > 500.0s | 0 | 0 | 0 |
| VertiFormer | Continuous (MPPI) | 0.7s | 40 | 0 | 0 |
| Hydra-NoVQ | Continuous (MPPI) | 1.2s | 20 | 10 | 10 |
| HYDRA (Ours) | Discrete (DLP) | 0.9s | 100 | 80 | 80 |
HYDRA achieves the "Goldilocks" zone—real-time edge performance, overcoming NWM's catastrophic latency and VertiFormer's sample inefficiency.
HYDRA's generative head produces high-fidelity predictions of future states. We demonstrate this in two ways: teacher-forced generation, which shows the model's rendering quality and adherence to the given ground-truth actions, and action head generation, which shows the model's ability to create coherent long-term rollouts on its own. Notice how action head rollouts that lead to collisions (out-of-distribution states) result in visual "hallucinations," an uncertainty signal we use for safe planning.
Model renders frames conditioned on ground-truth actions from the dataset, showing high-fidelity reconstruction.
Model predicts its own actions and generates future frames. The resulting "hallucination" upon collision is used as a cost for planning.
A set of videos produced by the image generation head.
We visualize the planning process by showing the robot's view alongside an RViz visualization of the plans being evaluated. This highlights how HYDRA selects a lowest cost path and executes it with a controller.
Robot view with RViz showing evaluated trajectories.
Execution of a planned trajectory (red) in a cluttered hallway. Yellow is uncond, purple is cond, and green is the planned path.
Sampling-based path following demonstration.
Real-world deployments across complex environments. HYDRA handles obstacles and narrow corridors.
Multiple image generation through time with one forward pass. Top row is the ground truth, bottom is the generated image.
No model is perfect. To foster transparent research, we highlight scenarios where HYDRA's performance degrades. Please see the paper's limitations section and failure analysis section in the Appendix.
Failure: Trapped in a tight space by not turning in time. The planner did not converge to the goal in time before a sharp right turn.
Failure: Vanishing Humans in Social Scenarios. The limited cardinality of the visual codebook causes degradation in social scenarios where there are lots of visual changes caused by humans who only occupy a small number of pixels.
@article{nazeri2026hydra,
title={Hydra: A Navigation World Action Model with Discrete Latent Planning and Continuous Flow-Matching Execution},
author={Nazeri, Mohammad and Card, Alexandyr and Huber, Samira and Pokhrel, Anuj and Wang, Yujun and Hammele, Ruben and Song, Daeun and Pirk, S{\"o}ren and Xiao, Xuesu},
journal={arXiv preprint arXiv:2608.28995},
year={2026}
}