Recovery-Guided Model-Based Diffusion for Safety-Critical Legged Locomotion

Lei Zheng, Shivam Sood, Mithun Vanniasinghe, Guillaume Sartoretti, Armin LedererarXiv 2026 · arXiv

TL;DR

This paper introduces a framework that gives short-horizon robot motion planners "long-term foresight" by using a learned recovery critic. By combining model-based diffusion with a neural network that predicts future safety, the authors achieved zero collisions in high-speed legged locomotion tests while using significantly less computation than traditional long-horizon planners.

What Problem Does This Paper Address?

High-speed legged robots (like the Unitree Go2) operate in a "safety-critical" regime. To avoid obstacles, they need to plan their movements several steps ahead. However, robot designers face a classic trade-off:

  1. Short prediction horizons are computationally cheap and allow for fast updates, but they are "blind" to dangers just beyond their sight (e.g., a robot might run into a corner it can't turn out of).
  2. Long prediction horizons provide better safety foresight but are computationally expensive, often making them too slow for real-time control in complex, high-dimensional systems.

Existing methods like Model-Based Diffusion (MBD) are great at exploring different ways to move, but they struggle to maintain strict safety constraints over short distances without "falling off a cliff" once the planned trajectory ends.

Methods

The authors propose Recovery-Guided Model-Based Diffusion (RG-MBD), which consists of three main components:

  • Terminal Recovery Critic: Instead of planning 50 steps ahead, the robot plans only 10 steps. At the end of those 10 steps, a neural network (the critic) evaluates the state. This critic is trained offline to estimate the "worst-case future clearance" if the robot were to follow a dedicated recovery policy. Essentially, it tells the planner: "If you end up here, can you still stay safe later?"
  • Feasibility-Guided Refinement: During the diffusion process (where the planner iteratively improves a batch of candidate trajectories), the system uses a "feasible-first" weighting. It prioritizes trajectories that satisfy both immediate physical constraints and the terminal recovery criteria.
  • Safe-Elite Retention: To prevent the stochastic sampling process from accidentally discarding all safe options (a common issue called "sampling-support loss"), the algorithm explicitly keeps a set of "elite" feasible trajectories to seed the next round of refinement.

Datasets

The study primarily uses synthetic data generated within the MuJoCo MJX physics engine:

  • Recovery-Rollout Dataset: To train the critic, the authors generated 2,048 recovery trajectories using a long-horizon (50-step) controller.
  • Evaluation Benchmark: The framework was tested on a Unitree Go2 robot across 50 paired randomized obstacle-avoidance trials. These trials involved the robot running at 1.5 m/s toward obstacles of varying positions and sizes.

Key Insights and Ideas

  • Recoverability as a Constraint: The "aha moment" here is treating the ability to recover as a terminal constraint. This effectively extends the safety horizon to infinity (under a specific policy) without the overhead of simulating every step.
  • Mitigating Support Loss: In constrained optimization via diffusion, the "feasible" region of the search space can be very small. By using "safe-elite retention," the authors ensure the planner doesn't "forget" a safe path it found in an earlier iteration.
  • Model-Based vs. Data-Driven: Unlike standard Diffusion Policies that learn from human demonstrations, this is Model-Based. It uses the robot's actual physics model to "roll out" trajectories, ensuring that the movements are physically possible for the hardware.

Key Results

The proposed 10-step RG-MBD planner was compared against several baselines, including DIAL (a state-of-the-art diffusion planner) at various horizons:

  • Safety: RG-MBD achieved 0% collision rate and 98% strict-clearance success (maintaining at least 2cm of space). In contrast, even a 50-step DIAL planner had a 58% collision rate.
  • Efficiency: RG-MBD required a mean planning latency of 100.1 ms, significantly faster than the 273.3 ms required by the 50-step DIAL baseline.
  • Reliability: In ablation studies, the authors showed that without the "safe-elite" retention, the planner often failed because it would lose track of the few feasible trajectories in the high-dimensional search space.

Limitations and Future Work

  • Fixed Recovery Policy: The critic is currently trained on a single, fixed recovery policy. If the environment changes in a way that the recovery policy can't handle, the critic's advice might become invalid.
  • Simulation to Reality: While MuJoCo MJX is highly accurate, the paper does not yet demonstrate these results on physical hardware, where sensor noise and model discrepancies (Sim-to-Real gap) are more prevalent.
  • Terrain Complexity: The current work focuses on flat ground with box-like obstacles. Future work aims to extend this to uneven, rugged terrain where the physics of contact becomes even more complex.

Summary updated

Abstract

Safety-critical high-speed legged locomotion requires motion planners to balance safety foresight with online computational efficiency. Although short prediction horizons reduce rollout costs, they provide limited safety foresight, and strict safety constraints severely restrict the feasible trajectory space. We propose a recovery-guided model-based diffusion (MBD) framework for safety-critical legged locomotion. It incorporates a learned terminal recovery critic into short-horizon trajectory optimization. Specifically, the critic approximates the worst-case future clearance margin along rollouts of a designated recovery policy. We introduce a feasibility-guided refinement strategy that combines within-horizon constraints and terminal recoverability at each sampling stage, while retaining feasible elites to mitigate sampling-support loss. Across 50 paired randomized Unitree Go2 obstacle-avoidance trials, our approach achieves zero collisions and 98% strict-clearance success, while requiring substantially less computation than long-horizon MBD baselines. A project page is available at https://zack4417.github.io/rg-mbd/.
BibTeX
@misc{zheng2026recovery,
  author = {Lei Zheng and Shivam Sood and Mithun Vanniasinghe and Guillaume Sartoretti and Armin Lederer},
  title = {{Recovery-Guided Model-Based Diffusion for Safety-Critical Legged Locomotion}},
  year = {2026},
  eprint = {2611.19775},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19775},
}
Figure