[Paper Review] Fastron: An Online Learning-Based Model and Active Learning Strategy for Proxy Collision Detection

N. R. Das, Naman Gupta+1 more|arXiv (Cornell University)|Sep 7, 2017

Machine Learning and Algorithms4 citations

TL;DR

Fastron proposes an online learning-based proxy collision detection model using a modified kernel perceptron and active learning to accelerate C-space collision checks. It achieves up to 8× speedup over high-precision kinematic collision detection without GPU or parallelization, maintaining high recall and low false positive rates in dynamic environments with iterative model updates.

ABSTRACT

We introduce the Fastron, a configuration space (C-space) model to be used as a proxy to kinematic-based collision detection. The Fastron allows iterative updates to account for a changing environment through a combination of a novel formulation of the kernel perceptron learning algorithm and an active learning strategy. Our simulations on a 7 degree-of-freedom arm indicate that proxy collision checks may be performed at least 2 times faster than an efficient polyhedral collision checker and at least 8 times faster than an efficient high-precision collision checker. The Fastron model provides conservative collision status predictions by padding C-space obstacles, and proxy collision checking time does not scale poorly as the number of workspace obstacles increases. All results were achieved without GPU acceleration or parallel computing.

Motivation & Objective

• To reduce the computational burden of kinematic-based collision detection (KCD) in robot motion planning and iterative tasks.
• To enable real-time adaptation of C-space obstacle representations in response to changing environments without requiring prior knowledge of obstacle geometries or trajectories.
• To develop a lightweight, incremental learning model that maintains high accuracy while minimizing costly KCD queries through active learning.
• To demonstrate that proxy collision detection can significantly accelerate sampling-based motion planners like RRT without sacrificing safety or accuracy.
• To provide a scalable, hardware-agnostic solution for collision detection that complements rather than replaces KCD, using KCD only for critical model updates.

Proposed method

• Fastron uses a modified kernel perceptron algorithm to learn a decision boundary in configuration space (C-space) that separates free (C_free) and collision (C_obs) configurations.
• The model supports both addition and removal of support points, enabling iterative updates to the decision boundary as obstacles move or are introduced.
• An active learning strategy selects the most informative configurations for querying the expensive KCD oracle, minimizing the number of required KCD evaluations.
• The algorithm uses periodic snapshots of obstacle positions and shapes to update the model, allowing it to adapt to dynamic environments without retraining from scratch.
• C-space obstacles are padded conservatively to ensure the proxy model does not miss actual collisions, maintaining safety in motion planning.
• The method operates entirely in CPU-based environments without GPU acceleration or parallelization, relying on efficient kernel computations and sparse support point representation.

Experimental results

Research questions

• RQ1 Can a learning-based proxy model reduce collision check time in C-space while maintaining high accuracy and safety in dynamic environments?
• RQ2 How effectively can an online kernel perceptron with active learning reduce the number of required KCD queries during model updates?
• RQ3 To what extent does the Fastron model scale with increasing numbers of workspace obstacles and changing configurations?
• RQ4 How does the Fastron compare in speed and accuracy to high-precision and high-speed KCD methods in real-time motion planning scenarios?
• RQ5 Can the Fastron model be incrementally updated to reflect environmental changes without retraining, and how does this affect performance?

Key findings

• Fastron achieves at least 2× speedup over an efficient polyhedral collision checker and 8× speedup over a high-precision collision checker in 7 DOF manipulator simulations.
• With N=4000 support points, Fastron maintains a recall of over 90% and a false positive rate below 15% when using FCL for KCD, even as obstacle count increases.
• In RRT motion planning, the Fastron-based collision checking stage was 3.7× faster than KCD-RRT when using FCL, and 1.6× faster when using GJK.
• Model update time increased with the relabeling allowance A, but remained under 30 ms for typical configurations, with update times of 123.3 ms at N=8000 and A=0.3N.
• Recall decreased slightly with larger datasets (e.g., 90.6% at N=8000) but remained above 90%, while false positive rate improved with larger N due to better boundary coverage.
• The method operates efficiently on CPU without GPU acceleration or parallelization, demonstrating hardware-agnostic performance gains.