Abstract
Diffusion models have emerged as a powerful paradigm for robotic motion planning, yet they frequently struggle to satisfy the hard, implicit safety constraints required for complex locomotion. We introduce a framework that decouples plan generation from constraint evaluation using Learned Viability Filters (VFs). By utilizing fast, composable Q-functions to evaluate a batch of diffusion-generated trajectories, our approach identifies the plans most likely to ensure long-term stability. In 3D humanoid locomotion tasks, this method achieves significantly higher success rates and faster inference times than traditional classifier-based guidance, demonstrating a robust path forward for safe, real-time robotic planning.
Introduction
Generative modeling, particularly diffusion, has transformed robotic control by allowing agents to learn complex multi-modal behaviors from large-scale trajectory data. Models like Diffuser and Diffusion Policy can generate plausible future states and actions by denoising random noise into structured plans. However, these models possess a critical weakness: they struggle with hard constraints.
In high-stakes environments—such as a humanoid robot navigating a narrow platform or clearing a hurdle—the margin between success and failure is razor-thin. A diffusion model trained on stochastic data may capture the general "shape" of a successful movement but fail to respect the strict boundaries of physical safety, leading to catastrophic falls. Traditionally, researchers have addressed this using classifier guidance, which uses gradients to steer the diffusion process toward safe regions. However, guidance is computationally expensive and often difficult to tune. This paper proposes a more efficient alternative: generating a diverse set of candidate plans and filtering them through a specialized viability evaluator.
The Viability Filtering Framework
The proposed architecture splits the planning process into two distinct components: a Diffusion Planner and a Viability Filter (VF).
1. The Diffusion Planner
The system utilizes a 1D U-Net diffusion model trained on successful motion trajectories. At each time step, the planner proposes a batch of $N$ short-term candidate plans (e.g., sequences of four footsteps). Because the diffusion model is trained on a broad distribution of successful data, these proposals are inherently diverse and physically plausible, though not guaranteed to be safe in the current specific environment.
2. The Viability Filter (VF)
The VF is a neural network trained as a Q-function. It takes the current state and a proposed plan as input and outputs the expected discounted cumulative reward—effectively the probability that the plan will keep the robot "alive" without falling.
3. Inference and Execution
During inference, the system generates $N$ candidate plans in parallel. The VF scores each plan, and the system selects the highest-scoring trajectory. The robot executes the first action of this optimal plan and then repeats the process (replanning) at the next time step. This hybrid approach combines the generative power of diffusion with the discriminative precision of value-based reinforcement learning.
Key Innovations
Online Training and Distribution Shifts
While VFs can be trained on offline datasets, the authors found that they perform best when trained online using an experience replay buffer of the diffusion model's actual outputs. This allows the VF to learn the specific failure modes of the planner, correcting for distribution shifts between the original training data and the generated plans.
Asymmetric Conditioning
A significant advantage of this decoupling is asymmetric conditioning. The diffusion model can remain "blind" to certain environmental features (like the exact location of a hurdle), focusing solely on the mechanics of walking. Meanwhile, the VF is fed the specific locations of obstacles. This allows the VF to steer the robot around hazards without requiring the underlying diffusion model to be retrained for every new environment.
Compositionality
Because VFs output expected viability, multiple task-specific filters can be composed zero-shot. For example, a filter for "avoiding obstacles" and a filter for "stepping over hurdles" can be multiplied together at inference time. This allows the robot to handle complex, multi-constraint environments by combining simple, specialized modules.
Experimental Results
The framework was evaluated on a simulated 3D humanoid using the ALLSTEPS framework for low-level control. The tasks included navigating high platforms, avoiding obstacles, and clearing hurdles.
Performance on Hard Constraints
On the "Platform Task," where the robot must navigate heights up to 65 cm (exceeding the training distribution of the low-level controller), the results were stark:
- Diffusion Only: 21% success rate.
- Diffusion + Offline VF: 39% success rate.
- Diffusion + Online VF: 68% success rate.
In obstacle avoidance tasks involving 1.5m radius cylinders, the baseline diffusion model failed 100% of the time. While classifier-based guidance improved this to 63%, the Online VF achieved a 99% success rate.
Efficiency and Speed
The VF approach is significantly more efficient than gradient-based guidance. Evaluating a batch of 100 plans with the VF takes approximately 0.105 seconds. In contrast, generating a single plan using classifier guidance takes 0.387 seconds. Thus, the VF approach is nearly four times faster while evaluating 100 times more candidate trajectories.
Conclusion and Future Work
By decoupling plan generation from viability evaluation, this research provides a robust solution to the safety limitations of diffusion-based planners. The use of Learned Viability Filters allows for fast, composable, and highly accurate planning in constrained environments.
However, the method is currently limited by the diffusion model's proposal distribution; the VF can only select the best plan from those provided. If the diffusion model never proposes a viable path (e.g., failing to suggest a necessary turn), the filter cannot intervene. Future work will explore iterative loops where the VF's discoveries are used to fine-tune the diffusion model, potentially enabling more sophisticated exploration and even higher success rates in novel environments.
Summary updated