Just in time Informed Trees: Manipulability-Aware Asymptotically Optimized Motion Planning

In high-dimensional robotic path planning, traditional sampling-based methods often struggle to efficiently identify both feasible and optimal paths in complex, multi-obstacle environments. This challenge is intensified in robotic manipulators, where the risk of kinematic singularities and self-collisions further complicates motion efficiency and safety. To address these issues, we introduce the Just-in-Time Informed Trees (JIT*) algorithm, an enhancement over Effort Informed Trees (EIT*), designed to improve path planning through two core modules: the Just-in-Time module and the Motion Performance module. The Just-in-Time module includes “Just-in-Time Edge,” which dynamically refines edge connectivity, and “Just-in-Time Sample,” which adjusts sampling density in bottleneck areas to enable faster initial path discovery. The Motion Performance module balances manipulability and trajectory cost through dynamic switching, optimizing motion control while reducing the risk of singularities. Comparative analysis shows that JIT* consistently outperforms traditional sampling-based planners across $\mathbb{R}^4$ to $\mathbb{R}^{16}$ dimensions. Its effectiveness is further demonstrated in single-arm and dual-arm manipulation tasks, with experimental results available in a video at https://youtu.be/nL1BMHpMR7c.

💡 Research Summary

The paper introduces Just‑in‑Time Informed Trees (JIT*), an anytime optimal motion‑planning algorithm that builds on the earlier Effort‑Informed Trees (EIT*) framework. JIT* targets two persistent challenges in high‑dimensional robotic manipulators: (1) the difficulty of quickly finding an initial feasible path and converging to an optimal solution, especially in cluttered or narrow‑passage environments, and (2) the neglect of manipulability and singularity avoidance in most sampling‑based planners, which can lead to infeasible joint velocities or self‑collisions.

JIT* consists of two complementary modules. The Just‑in‑Time module contains (a) Just‑in‑Time Edge, which dynamically expands and rewires edges beyond the parent‑only restriction of classic RRT* by exploiting a reverse lazy search tree. This edge‑expansion propagates low‑cost information throughout the tree, allowing rapid reduction of path length in bottleneck regions. (b) Just‑in‑Time Sample, which monitors collision‑heavy zones and adaptively increases sampling density there, rather than discarding whole batches as in EIT*. This localized resampling preserves exploration efficiency while focusing computational effort where it matters most.

The Motion Performance module augments the traditional cost function with a manipulability‑aware term. The total cost S_total is split into two sub‑objectives: s₁_sub, an admissible heuristic composed of start‑to‑goal (ĥ), edge (ĉ) and start‑to‑node (ĝ) costs, which drives the primary search for a low‑cost trajectory; and s₂_sub, a tie‑breaker that incorporates a manipulability measure D_tanh and an effort estimate ē, penalising configurations near kinematic singularities or self‑collision. A scaling factor α balances the units of the two terms, enabling the planner to switch dynamically between pure distance optimisation and manipulability‑aware optimisation as the search progresses.

Algorithmically, JIT* maintains a reverse lazy tree (R‑Tree) and a forward feasibility tree (F‑Tree). After initializing the start state and a manipulability‑optimised goal configuration (x*₍goal₎), the reverse search quickly identifies a low‑cost backbone using the S_JIT*^R cost. The forward tree then attempts to improve this solution; if a new edge can reduce the total cost, it is incorporated, otherwise Just‑in‑Time Sample adds points in critical regions and the reverse search restarts. The loop continues until a user‑defined time or iteration limit is reached, after which a pruning step extracts the optimal path π*.

Experimental evaluation spans synthetic problems in ℝ⁴–ℝ¹⁶, a 6‑DOF single‑arm task, and a 12‑DOF dual‑arm manipulation scenario. JIT* is benchmarked against RRT*, RRT*#, BIT*, EIT*, FDIT*, and FCIT*. Results show that JIT* discovers an initial feasible path 30‑40 % faster and yields final trajectories that are 15‑25 % lower in total cost. Crucially, in situations where the shortest‑geodesic path would pass near singularities, traditional planners either fail or produce infeasible joint velocities, whereas JIT*’s Motion Performance module steers the search toward safer, manipulability‑rich configurations without sacrificing overall optimality.

In summary, JIT* advances high‑dimensional manipulation planning by (1) leveraging dynamic edge rewiring for global cost propagation, (2) employing adaptive, bottleneck‑focused sampling, and (3) integrating a manipulability‑aware cost term that dynamically balances path length against kinematic safety. The authors suggest future work on extending JIT* to dynamic environments with moving obstacles and to coordinated multi‑robot systems, as well as tighter integration with real‑time control loops.