Fast GPU Evaluation of Differentiable Signed Distance Fields for Robotics via Tensor Decompositions
| Type of publication: | Journal paper |
| Citation: | Govoni_RAS_2026 |
| Publication status: | Published |
| Journal: | Robotics and Autonomous Systems |
| Year: | 2026 |
| DOI: | 10.1016/j.robot.2026.105394 |
| Abstract: | Safe and efficient collision avoidance is essential for robots operating in dynamic and cluttered environments. We present a task-priority control framework that embeds signed distance fields (SDFs) directly into the control loop, enabling smooth and reactive avoidance of both environmental and self-collisions. Robot links are represented with Bernstein polynomial-based distance fields, which provide continuous geometry models and closed-form gradients for defining repulsive actions. These avoidance behaviors are activated seamlessly within the task hierarchy and executed in real time through a GPU-accelerated implementation. The framework is validated on fixed base and mobile manipulators exposed to dynamic obstacles sensed with depth cameras and laser scanners. Results show consistent improvements in responsiveness, computational efficiency and motion smoothness compared to conventional optimization-based approaches, demonstrating the effectiveness of integrating SDFs into task-priority control for robust robot motion in unstructured environments. |
| Main Research Program: | Human-AI Teaming |
| Keywords: | robot learning, robot manipulation, robotics |
| Projects: |
Idiap |
| Authors: | |
| Added by: | [UNK] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
