acceptodds
Under review as a conference paper at ICLR 2027

Hydra: A Navigation World Action Model with Discrete Latent Planning and Continuous Flow-Matching Execution

Abstract

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 entirely 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. Evaluated on two physical robotic platforms, Hydra outperforms state-of-the-art navigation world models in goal-directed planning, while matching or exceeding the closed-loop execution capabilities of leading reactive navigation policies

Primary area:
applications to robotics, autonomy, planning
Keywords:
world action model, navigation, planning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HyvZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Hydra, a multimodal world action model for navigation that uses a single latent space for multimodal information, unifies the future predictions and policy planner in latent rollouts instead of separate modules. Then, the planning is formulated as search on discrete latent space which guided by a multi-objective Kinematic-Perceptual Cost. Experiments show the performance of Hydra and the comparison to navigation policy baselines.

Critical strengths

The author provides physical evaluations, which give more useful evidence than the offline benchmarks. The hardware deployment and experimental design are actually helpful for understanding the effectiveness of the system design than the visual metrics. The appendix also discusses the misalignment between the actual physical evaluation results and the benchmarks. Also, the results show the strong performance of Hydra, achieving a high success rate in planning scenarios.

Critical weaknesses
  1. The ablation is not a sufficiently controlled study, the Hydra-NoVQ removes the discretization but also uses different cost terms and search configurations. Thus, this is not enough to claim that the VQ discretization is an improvement in the system design.
  2. In the experiments, VertiFormer and Hydra-NoVQ only get the geometric goal and image goal, without an obstacle term, so the low performance in obstacle scenes is expected.
Questions for authors
  1. Do the NWM 0% results come from the performance (quality) or from the latency?
  2. Can the author add a more controlled ablation study for Hydra-NoVQ?
Reviewer_L9HKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Hydra, a navigation world model that combines discrete latent planning with continuous Flow Matching. The method uses VQ codebooks to represent visual states and motion intents, enabling trajectory planning and evaluation in latent space. Experiments on two real robotic platforms demonstrate promising navigation performance and planning efficiency.

Critical strengths

1.Interesting Design: The integration of discrete latent planning and continuous action generation provides an interesting approach to robot navigation. 2.Good Planning Efficiency: Hydra significantly reduces planning time compared with NWM while achieving good success rates in obstacle avoidance tasks. 3.Real-World Evaluation: The paper includes experiments on both Jackal and Spot robots, along with several useful ablation studies.

Critical weaknesses

1.Limited Novelty and Ablation Studies: Discrete latent planning and latent-space evaluation have been explored in previous work. Moreover, Table 2 changes the search space, cost functions, and search settings simultaneously, making it difficult to determine the contribution of VQ discretization itself. 2.Limited Validation of Safety Evaluation: The proposed KPC uses predictive entropy and VQ quantization error to estimate collision risks. Although Table 8 shows their usefulness, the failure cases in Appendix B suggest that these signals may confuse visual complexity with actual obstacles. Further quantitative evaluation is needed. 3.Limited Long-Distance Performance: Hydra achieves only 40% success on the Long Distance task in planning mode (Table 3), despite performing well in local path following. This raises concerns about the effectiveness of the proposed planning method in longer navigation tasks.

Questions for authors

1.Additional Ablation Experiments: Could the authors compare Hydra and Hydra-NoVQ using the same search budget and shared goal cost functions? The authors could then add the remaining KPC terms to Hydra to better demonstrate the contribution of discrete planning. 2.Safety Evaluation: Could the authors evaluate predictive entropy and VQ quantization error on scenarios involving obstacles, complex but traversable textures, and pedestrians? Reporting collision risk detection accuracy and false positive rates would help assess the reliability of these signals. 3.Long-Distance Evaluation: Could the authors compare different numbers of search iterations (e.g., 1, 2, and 3) on the Long Distance task? Reporting success rates and planning times would help explain the current performance limitations.

Reviewer_bifJOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Hydra combines multimodal world modeling, discrete latent planning, and continuous Flow Matching for robot navigation. It ranks candidate trajectories using goal-related, prior, and uncertainty-based costs without pixel-level decoding, then generates continuous execution trajectories. Experiments on two robot platforms demonstrate substantially lower planning latency than the evaluated NWM baseline and promising local navigation performance.

Critical strengths

Practical design: Discrete search and deferred generative decoding address an important computational bottleneck in world-model-based planning.

Interesting latent evaluation: Predictive entropy and quantization error provide inexpensive risk signals, with component ablations supporting their usefulness.

Physical validation: Experiments on two robot platforms and explicit failure analyses strengthen the empirical contribution.

Critical weaknesses

Insufficient isolation of discretization: The Hydra-NoVQ comparison also changes available cost terms, search strategy, and iteration count. Consequently, the results do not cleanly establish that discretization itself drives the gains.

Limited support for safety claims: Entropy and quantization error are imperfect collision proxies. The reported texture-related false positives and missing pedestrians warrant broader quantitative evaluation and more cautious safety claims.

Efficiency claims need qualification: The reported >500× speedup concerns a specific NWM configuration. Inference runs on an external A5000 workstation, and end-to-end control frequency and runtime breakdowns are insufficiently characterized.

Limited and mixed evaluation: Each scenario contains only ten trials, and full planning achieves 40% success on the long-distance task. Comparisons with reactive policies also use different goal representations, complicating conclusions about overall superiority.

Questions for authors

Please see the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.