Sampling-based Model Predictive Control Leveraging Parallelizable Physics Simulations

Abstract: We present a method for sampling-based model predictive control that makes use of a generic physics simulator as the dynamical model. In particular, we propose a Model Predictive Path Integral controller (MPPI), that uses the GPU-parallelizable IsaacGym simulator to compute the forward dynamics of a problem. By doing so, we eliminate the need for explicit encoding of robot dynamics and contacts with objects for MPPI. Since no explicit dynamic modeling is required, our method is easily extendable to different objects and robots and allows one to solve complex navigation and contact-rich tasks. We demonstrate the effectiveness of this method in several simulated and real-world settings, among which mobile navigation with collision avoidance, non-prehensile manipulation, and whole-body control for high-dimensional configuration spaces. This method is a powerful and accessible open-source tool to solve a large variety of contact-rich motion planning tasks.

• The paper introduces a sampling-based MPC using IsaacGym, eliminating the need for explicit dynamic modeling in contact-rich environments.
• It employs Halton Splines for efficient stochastic sampling, improving exploration and convergence in mobile navigation and whole-body control.
• Experimental results demonstrate competitive motion planning, robust collision avoidance, and effective real-world deployment.

Sampling-based Model Predictive Control Leveraging Parallelizable Physics Simulations

Introduction

The paper introduces a sampling-based model predictive control (MPC) method utilizing the IsaacGym physics simulator for dynamic modeling. By leveraging IsaacGym's GPU-parallelizable simulation capabilities, the proposed Model Predictive Path Integral (MPPI) controller can efficiently perform rollouts to solve complex contact-rich tasks without requiring explicit dynamic modeling. This flexibility enables application across various scenarios, including mobile navigation and high-dimensional whole-body control, making the approach robust and versatile for real-world robotic tasks.

Figure 1: Scheme of the proposed method using IsaacGym as the dynamic model for MPPI.

Model Predictive Path Integral (MPPI) Framework

MPPI is based on sampling stochastic optimal control inputs for discrete-time systems to generate state trajectories. These are evaluated against a cost function to optimize the control sequence. Traditionally, MPPI requires accurate dynamic modeling, which can be challenging for contact-rich and complex environments. The proposed framework eliminates this necessity by using IsaacGym to simulate dynamics, thereby reducing computational overhead and simplifying the process of handling non-convex dynamics.

Figure 2: Examples for pure motion planning benchmark setups. Left: point robot with 3 DOF. Right: manipulator with 7 DOF.

Algorithm Implementation

The algorithm initializes with an input sequence, samples noise using Halton Splines, and simulates state trajectories in parallel using IsaacGym. Costs are computed for each trajectory, and importance sampling weights are used to obtain the optimal control input sequence applied to the robot. The method relies on dynamically updating weights and exploring input space efficiently via Halton Splines rather than Gaussian sampling, allowing improved convergence and exploration.

Collision Checking and Uncertainty Management

Collision checking is performed using IsaacGym's contact forces tensor, enabling continuous assessment during task execution without additional convexification. Domain randomization within IsaacGym addresses model uncertainties by varying object properties, maintaining robustness across diverse environments.

Experimental Results

Motion Planning and Collision Avoidance

Experiments reveal comparable performance with state-of-the-art methods in motion planning without environmental interactions but with faster goal convergence due to precise representation of robot shapes and dynamics. Our method incurs higher computational times due to complex simulations, yet retains competitive advantages in contact-rich scenarios.

Whole-body Control

The proposed system seamlessly handles complex tasks such as mobile manipulation involving multiple DOFs without needing to segment or pre-plan motion sequences. Demonstrations include relocating objects using a mobile manipulator, showcasing fluid, whole-body movement directly derived from the cost function.

Figure 3: Whole-body motion of a mobile manipulator moving a cube from initial to desired location.

Non-prehensile Manipulation

The system effectively addresses pushing tasks with both robotic arms and mobile bases, outperforming baseline methods which require learning dynamics or limiting sampled trajectories. Experiments demonstrate smooth manipulation of both boxes and spheres by various robot platforms, highlighting the versatility of the approach.

Figure 4: Rolling ball non-prehensile pushing. Goal: Ball placement between two obstacles.

Real-world Deployment

Real-world trials demonstrate comparable performance to simulations in non-prehensile manipulation tasks, reinforcing the applicability of the method in practical settings. Disturbances during execution are compensated by the control strategy, ensuring robust interaction with dynamic environments.

Figure 5: Qualitative real-world experiments with disturbances. The behavior can be seen in the accompanying video.

Discussion

The proposed method is computationally intensive, especially with extended horizons, suggesting future integration with global planning strategies to improve sampling efficiency. Tuning for control algorithms remains a challenge; autotuning techniques could mitigate this issue. Further enhancement through sensor integration may facilitate collision avoidance in complex setups.

Conclusion

The integration of MPPI with IsaacGym presents a powerful framework for robotic control in contact-rich environments, avoiding the complexities of explicit dynamic modeling and contact engineering. The approach is validated across diverse tasks, demonstrating faster execution and improved accuracy while offering a versatile open-source solution adaptable to various robotic applications.