Efficient Motion Planning for Manipulators with Control Barrier Function-Induced Neural Controller

Mingxin Yu, Chenning Yu, M.-Mahdi Naddaf-Sh, Devesh Upadhyay, Sicun Gao, Chuchu Fan

Abstract

Motion planning for robotic manipulators in cluttered environments remains a significant challenge due to high sampling complexity and the computational cost of collision checking. This paper introduces CBF-INC-RRT, a motion planning framework that integrates the global search capabilities of Rapidly-exploring Random Trees (RRT) with the local safety guarantees of learned Control Barrier Functions (CBFs). By replacing the standard linear steering in RRT with a neural-network-based controller that understands complex robot geometry and raw LiDAR data, the proposed method achieves higher success rates with fewer explored nodes. Experimental results on 4-DoF and 7-DoF manipulators demonstrate that the framework reduces over-conservatism and improves planning efficiency in hard-to-navigate spaces.

Introduction

Traditional sampling-based methods, such as RRT, are the workhorses of robotic motion planning. They operate by randomly sampling the configuration space and attempting to connect these samples to a growing tree. However, in crowded or narrow environments, these methods suffer from two primary bottlenecks: high sampling complexity, where most random samples result in collisions, and expensive collision checking, which requires constant geometric validation of path segments.

While Control Barrier Functions (CBFs) offer a mathematically rigorous way to ensure real-time collision avoidance, they are traditionally hand-crafted using simplified geometric primitives like spheres or boxes. This simplification often leads to over-conservative behavior, where a robot may fail to navigate through a narrow gap it could physically fit through because its "safety bubble" is too large. This paper addresses these limitations by learning the safety landscape directly from data and embedding this knowledge into the RRT steering function.

Methodology

The authors propose a two-stage approach to create an obstacle-aware steering function that guides the RRT expansion more effectively.

1. Learning the Safety Landscape (CBF-INF)

The core of the framework is a neural network trained to represent a CBF, denoted as $h(q, o)$, which maps the robot's configuration ($q$) and its observations ($o$) to a safety value. Two variants were developed:

  • sCBF-INF: A state-based model using signed distance functions to obstacles.
  • oCBF-INF: An observation-based model that processes raw LiDAR point clouds through a PointNet architecture. This allows the system to operate directly from sensor data without requiring pre-defined 3D models of the environment.

To ensure the controller can maintain safety, the authors utilize numerical differentiation to calculate Lie derivatives. This approach allows the framework to treat the robot's kinematics as a "black box," making it easily adaptable to different robot models without the need for complex analytical derivations.

2. The CBF-Induced Controller (CBF-INC)

During the RRT expansion phase, the framework replaces the standard "blind" steering (moving in a straight line toward a sample) with a Quadratic Program (QP). This optimization finds a control signal that minimizes the distance to the target sample (using an LQR-based objective) while strictly adhering to the safety boundary defined by the learned neural CBF.

3. Integration with RRT

By using the CBF-INC as the local planner, the RRT tree becomes inherently obstacle-aware. The tree grows more efficiently toward the goal, as the steering function naturally avoids "dead ends" and navigates around obstacles during the expansion step itself, rather than relying solely on post-hoc collision checks.

Experimental Evaluation

The framework was evaluated using custom benchmarks for a 4-DoF Dobot Magician and a 7-DoF Franka Panda robot. The authors generated 1,000 "easy" and 1,000 "hard" test cases, where "hard" cases were defined by the time required for a state-of-the-art planner (BIT*) to find a solution.

Key Results

  • Improved Success Rates: On hard test cases, the state-based version (sCBF-INC-RRT) increased the success rate by 14% compared to vanilla RRT.
  • Search Efficiency: The method reduced the number of explored nodes by 30%, indicating that the planner finds valid paths much faster by making smarter local decisions.
  • LiDAR Robustness: When utilizing raw LiDAR input, the framework improved success rates by 10% over alternative neural steering controllers based on Reinforcement Learning or Imitation Learning.
  • Real-Time Performance: The controller demonstrated computational efficiency suitable for real-world deployment, with average signal computation times ranging from 0.15s to 0.32s.

Discussion and Conclusion

The primary insight of this work is that learned CBFs can capture the intricate, non-convex geometry of robot manipulators far more accurately than hand-crafted approximations. This reduction in over-conservatism allows robots to navigate tighter spaces with higher confidence. Furthermore, the integration of PointNet allows for a direct LiDAR-to-control pipeline, reducing the dependency on perfect environmental models.

Limitations and Future Work

Despite its advantages, the framework's performance is sensitive to LiDAR sensor noise, which can lead to safety violations if the point cloud is significantly degraded. Additionally, for robots with very high Degrees of Freedom, the computational cost of transforming point clouds into multiple link frames may become a bottleneck. Future research will focus on relaxing the assumptions regarding obstacle velocity to better handle fast-moving dynamic environments and improving the scalability of the point cloud processing pipeline.

Summary updated

Abstract

Sampling-based motion planning methods for manipulators in crowded environments often suffer from expensive collision checking and high sampling complexity, which make them difficult to use in real time. To address this issue, we propose a new generalizable control barrier function (CBF)based steering controller to reduce the number of samples needed in a sampling-based motion planner RRT. Our method combines the strength of CBF for real-time collision-avoidance control and RRT for long-horizon motion planning, by using CBF-induced neural controller (CBF-INC) to generate control signals that steer the system towards sampled configurations by RRT. CBF-INC is learned as Neural Networks and has two variants handling different inputs, respectively: state (signed distance) input and point-cloud input from LiDAR. In the latter case, we also study two different settings: fully and partially observed environmental information. Compared to manually crafted CBF which suffers from over-approximating robot geometry, CBF-INC can balance safety and goal-reaching better without being over-conservative. Given state-based input, our neural CBF-induced neural controller-enhanced RRT (CBFINC-RRT) can increase the success rate by 14% while reducing the number of nodes explored by 30%, compared with vanilla RRT on hard test cases. Given LiDAR input where vanilla RRT is not directly applicable, we demonstrate that our CBF-INCRRT can improve the success rate by 10%, compared with planning with other steering controllers. Our project page with supplementary material is at https://mit-realm.github.io/CBFINC-RRT-website/.
BibTeX
@inproceedings{yu2024efficient,
  author = {Mingxin Yu and Chenning Yu and M.-Mahdi Naddaf-Sh and Devesh Upadhyay and Sicun Gao and Chuchu Fan},
  title = {{Efficient Motion Planning for Manipulators with Control Barrier Function-Induced Neural Controller}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2024},
  url = {https://openreview.net/forum?id=eWYoOvOikJ},
}
Figure