STORM: An Integrated Framework for Fast Joint-Space Model-Predictive Control for Reactive Manipulation (2104.13542v2)

Abstract: Sampling-based model-predictive control (MPC) is a promising tool for feedback control of robots with complex, non-smooth dynamics, and cost functions. However, the computationally demanding nature of sampling-based MPC algorithms has been a key bottleneck in their application to high-dimensional robotic manipulation problems in the real world. Previous methods have addressed this issue by running MPC in the task space while relying on a low-level operational space controller for joint control. However, by not using the joint space of the robot in the MPC formulation, existing methods cannot directly account for non-task space related constraints such as avoiding joint limits, singular configurations, and link collisions. In this paper, we develop a system for fast, joint space sampling-based MPC for manipulators that is efficiently parallelized using GPUs. Our approach can handle task and joint space constraints while taking less than 8ms~(125Hz) to compute the next control command. Further, our method can tightly integrate perception into the control problem by utilizing learned cost functions from raw sensor data. We validate our approach by deploying it on a Franka Panda robot for a variety of dynamic manipulation tasks. We study the effect of different cost formulations and MPC parameters on the synthesized behavior and provide key insights that pave the way for the application of sampling-based MPC for manipulators in a principled manner. We also provide highly optimized, open-source code to be used by the wider robot learning and control community. Videos of experiments can be found at: https://sites.google.com/view/manipulation-mpc

• The paper presents a novel GPU-accelerated joint-space MPC framework that enables real-time reactive manipulation (<8ms latency) in robotic systems.
• It integrates learned cost functions and low-discrepancy sampling to manage complex constraints like joint limits, collision avoidance, and non-linear dynamics.
• Experimental validation on a Franka Panda robot demonstrates improved accuracy and smoother trajectories in dynamic target tracking and manipulation tasks.

Overview of "STORM: An Integrated Framework for Fast Joint-Space Model-Predictive Control for Reactive Manipulation"

The paper presents "STORM: An Integrated Framework for Fast Joint-Space Model-Predictive Control (MPC) for Reactive Manipulation," addressing the computational challenges posed by applying MPC to robotic manipulators. Typically, manipulating high-dimensional systems involves complex, non-smooth dynamics and constraints that are challenging to resolve in real-time. The authors propose a joint-space sampling-based MPC approach implemented in a highly parallelized manner using GPU acceleration to mitigate these challenges.

Utilizing the robot's joint space directly addresses potential issues neglected in task-space MPC, such as joint limits, singular configurations, and collision avoidance. This integration into STORM allows for consideration of both joint and task space constraints while achieving a computational latency of less than 8 milliseconds (125Hz), making real-time reactive manipulation feasible.

Key Contributions and Methodology

Several key techniques and methodologies underpin the STORM framework:

1. Sampling-Based MPC in Joint Space: The approach develops a joint-space, sampling-based MPC algorithm that exploits modern GPU capabilities to increase computational speed. This advancement allows real-time control for manipulators with complex, non-linear cost functions.
2. Integration of Perception: STORM incorporates perception directly into the control process by using learned cost functions. This element is crucial for handling environmental interactions based on real-time sensor data, allowing dynamic task execution and reactive behavior adjustments.
3. Low-Discrepancy Sampling: To alleviate sample clustering issues of traditional Monte Carlo methods, the authors employ low-discrepancy sequences for sampling actions, enhancing the stability and performance of the control algorithm.
4. B-Spline and Policy Parameterization: B-spline interpolation for action smoothing, coupled with adaptive covariance parameterization along action dimensions, maintains trajectory smoothness and improves task accuracy.
5. Open-Source Implementation: The authors contribute an optimized, open-source codebase for the wider robotics community, promoting further advancements in robot learning and control.

Experimental Validation

The framework's empirical evaluation on a real-world Franka Panda robot demonstrates effectiveness across various tasks. These include tracking moving targets with perception-driven feedback and dynamic ball balancing on a tray. Notably, STORM successfully handles challenging task constraints, such as maintaining end-effector orientation and performing collision avoidance.

Quantitative assessments reveal that increased particle counts lead to better control accuracy, and leveraging low-discrepancy, mixed sampling strategies improves trajectory smoothness and task success rates. The benchmark comparisons underscore the computational efficiency and practical applicability of STORM relative to existing control methods.

Implications and Future Directions

The STORM framework sets a new precedent for real-time joint-space control in robotic manipulation. Its integration of fast, parallelizable control strategies with perceptual feedback aligns well with the increasing demand for intelligent robots capable of autonomous and interactive operation.

Future developments could address potential limitations such as handling higher speed tasks where kinematic models might introduce significant bias. Extending the framework to incorporate uncertainty directly, possibly through learning-based residual dynamics models or informed terminal value functions, could further enhance the robustness of control.

In sum, this paper contributes significantly to the domain of robotic control, providing a novel, efficient framework capable of expanding the operational scope of robotic manipulation systems.