GRACE: Gradient-Free Robot Action Generation Beats Baseline Success Rates
New technique combines diffusion policies with MPPI to handle nondifferentiable safety constraints.
GRACE tackles a critical limitation of diffusion-based robot action generation: steering actions toward deployment-time constraints typically requires differentiable guidance costs. This excludes practical safety constraints like binary collision checks, joint limits, or black-box rollout costs. The method leverages the shared score-ascent structure of diffusion models and Model Predictive Path Integral (MPPI) control. At each reverse diffusion step, GRACE constructs a cost-conditioned guidance posterior and estimates its mean with a single MPPI update centered at the diffusion reverse mean. For differentiable costs, the approach recovers conventional gradient guidance under a first-order, matched-covariance approximation.
In simulation, GRACE outperformed diffusion-based and sampling-based baselines in success rates. On a real 7-degree-of-freedom manipulator, the method achieved 100% obstacle avoidance in every trial, while the unguided prior collided in every trial. The work demonstrates that gradient-free guidance using MPPI can effectively enforce nondifferentiable constraints during deployment. Code and experiment videos are available online.
- Guides pretrained diffusion policies using only forward cost evaluations, no gradients required for nondifferentiable constraints like binary collision checks or joint limits.
- Achieves 100% obstacle avoidance on a real 7-DoF manipulator, compared to the unguided prior colliding in every trial.
- Recovers conventional gradient guidance for differentiable costs as a special case under a matched-covariance approximation.
Why It Matters
Enables robots to safely navigate dynamic environments with practical constraints like collision checks and joint limits.