CoverLib: Classifiers-equipped Experience Library by Iterative Problem Distribution Coverage Maximization for Domain-tuned Motion Planning

TL;DR

CoverLib tackles the trade-off between planning speed and plannability by building a domain-tuned library where each experience is paired with a binary classifier that marks its adaptable problem-region. It employs an active, greedy procedure to maximize coverage in the problem-parameter space, training a base cost predictor and adjusting biases to meet a false-positive constraint, all in a domain-agnostic manner with respect to the adaptation algorithm. Across four diverse motion-planning domains, CoverLib demonstrates high success rates close to global planners, substantial speedups over purely global approaches, and scalability to high-dimensional problem spaces, while remaining compatible with both NLP-based and sampling-based adaptation methods. The work highlights nonlinear dimensionality reduction in adaptation regions and discusses limits, domain shift robustness, and practical guidance for deploying learning-augmented, library-based planners. These contributions advance domain-specific, high-plannability, fast planners suitable for real-world robotics and extended planning frameworks such as MPC and TAMP.

Abstract

Library-based methods are known to be very effective for fast motion planning by adapting an experience retrieved from a precomputed library. This article presents CoverLib, a principled approach for constructing and utilizing such a library. CoverLib iteratively adds an experience-classifier-pair to the library, where each classifier corresponds to an adaptable region of the experience within the problem space. This iterative process is an active procedure, as it selects the next experience based on its ability to effectively cover the uncovered region. During the query phase, these classifiers are utilized to select an experience that is expected to be adaptable for a given problem. Experimental results demonstrate that CoverLib effectively mitigates the trade-off between plannability and speed observed in global (e.g. sampling-based) and local (e.g. optimization-based) methods. As a result, it achieves both fast planning and high success rates over the problem domain. Moreover, due to its adaptation-algorithm-agnostic nature, CoverLib seamlessly integrates with various adaptation methods, including nonlinear programming-based and sampling-based algorithms.

Figures (24)

• Figure 1: Illustration of CoverLib's library and the problem (P-) space. (a) The experiences (paths) $\pi_{1:6}$ defined in a state space, e.g. configuration space or phase space. (b) The problem distribution $p(\theta)$ (contour plot) in P-space covered by the union of the estimated adaptable regions $\hat{F}_{1:6}$ (green regions) of the experiences. Note that although depicted in 2D here, the P-space is often very high (e.g. dozens or hundreds) dimensional as it encodes problem settings including obstacle environments, constraints, success condition specifications, etc. (c) Instantiation of three problems from the P-space.
• Figure 2: The proposed algorithm actively constructs a library to maximize coverage of $p(\theta)$. For visualization purposes, we assume a one-to-one correspondence between each problem and its solution (experience), thereby allowing experiences (e.g., best, cands, $\pi_k$) to be visualized within the P-space (note that this correspondence is not generally the case). Iterations $k=3$ and $k=5$ are highlighted, and steps 1 to 3 for each iteration are illustrated. Step 1 selects the next experience $\pi_k$, which is expected to yield the maximum coverage gain among the candidates. Step 2 trains the base term $\bar{f}_k$ via regression. Step 3 determines the bias terms $b_{1:k}$ through optimization to meet the FP rate constraint. During the optimization at step 3 of the $k$-th iteration, regions defined by the base term $\bar{f}_k$ (green dashed line) and by previously determined regions (black dashed lines) are either shrunk or expanded to the regions depicted respectively by green and black solid lines.
• Figure 3: Illustration of vector-encoder modeling.
• Figure 4: This figure visualizes the behavior of the CoverLib-aided planning with $(c_{\mathrm{max}}, \delta) = (5, 0.05)$ in Domain 1. The Yellow lines are the experiences in the library, while the red line is the selected experience used in adaptation. The blue line shows the result of adaptation. The purple circle and star denote the start and goal positions respectively. The blue line is not shown for failure cases as the result is not obtained. The label "fail(expected)" indicates cases where $\mathrm{min}_{i \in [K]} \hat{f}_i(\theta) > c_{\mathrm{max}}$, meaning failure is somewhat expected before adaptation. The label "fail(unexpected)" signifies that adaptation failed even though $\mathrm{min}_{i \in [K]} \hat{f}_i(\theta) \leq c_{\mathrm{max}}$. The overlaid "Global success" label for failed cases suggests that the problem is actually feasible, as evidenced by the success of Global.
• Figure 5: Benchmarking in terms of (a) GPE-TO, (b) Success rate, (c-d) Planning time for Domain 1. The values associated with the Global plot (e.g. 0.25, 0.5, ...) represent timeout values, which are depicted by short red bars in (c-d). The value associated with NNLib (e.g. $2^{20}$) indicates the library size. The values associated with CoverLib, e.g. (5, 0.05), are $(c_{\mathrm{max}}, \delta)$ pairs. In (c) and (d), the x-coordinate of each point in the scatter plot represents the problem index, while the y-coordinate represents the corresponding planning time. Data points for problems with failure are not plotted. The vertical pink lines indicate the minimum and maximum planning time of the scatter plot. The three horizontal pink lines along each scatter plot indicate the 50, 90, and 99 percentiles of the planning time. These percentile computations only consider successful cases. The approx. upper bound in (b) is computed as described in Section \ref{['sec:benchmarking_methodology']}.