Rapidly-exploring Random Tree Star (RRT*)

• Rapidly-exploring Random Tree Star (RRT*) is a sampling-based motion planner that guarantees probabilistic completeness and asymptotic optimality.
• Advanced variants, such as potential-guided and bidirectional RRT*, significantly reduce the number of samples and computation time in complex environments.
• Despite its strong theoretical foundation, RRT* may converge slowly in high-dimensional or cluttered spaces, prompting innovative acceleration strategies.

The Rapidly-exploring Random Tree Star (RRT*) algorithm is a sampling-based optimal motion planner that provides a probabilistically complete and asymptotically optimal solution for path planning problems in the presence of obstacles, regardless of their geometric complexity. RRT* incrementally constructs a tree that asymptotically converges to an optimal cost path as the number of samples increases. Despite its theoretical guarantees, RRT* often suffers from slow convergence in high-dimensional or highly cluttered environments, motivating a range of accelerations—including potential-guided sampling, bidirectional growth, and advanced rewiring heuristics—which build on the RRT* paradigm to achieve rapid, memory-efficient, and near-optimal behavior in practical settings.

1. Algorithmic Foundations of RRT*

RRT* operates in a continuous state space $X \subset \mathbb{R}^d$ with obstacle region $X_\mathrm{obs}$ and free space $X_\mathrm{free} = X \backslash X_\mathrm{obs}$. At each iteration, RRT* performs:

1. Sampling: Draws $$x_\mathrm{rand} \sim \mathrm{Uniform}(X_\mathrm{free})$$.
2. Near-Neighbor Search: Calculates a connection radius

$r_n = \gamma \left( \frac{\ln n}{n} \right)^{1/d}$

and finds all nodes $X_\mathrm{near}$ within $r_n$ of $x_\mathrm{rand}$; if none, uses the nearest neighbor.

1. Steering/Extension: For each $v \in X_\mathrm{near}$, computes a local path $\tau$ (typically a straight line), with cost $l = \mathrm{Cost}(\tau)$.
2. Collision Checking: Only considers $(v, \tau)$ pairs where $\tau \subset X_\mathrm{free}$.
3. Best-Parent Selection: Chooses the parent $v_\mathrm{parent}$ minimizing cumulative cost.
4. Insertion: Connects $x_\mathrm{rand}$ to $v_\mathrm{parent}$ and updates cost-to-come.
5. Rewiring: Reparents any $v_i \in X_\mathrm{near}$ if connection via $x_\mathrm{rand}$ is lower cost and feasible.

The cost metric is generally Euclidean path length:

$c(\tau) = \sum \| q_{k+1} - q_k \|$

Under regularity assumptions, RRT* is probabilistically complete (finds a solution if one exists as $n \to \infty$) and asymptotically optimal (cost converges almost surely to the optimum $c^*$) (Qureshi et al., 2017).

2. Limitations and Need for Acceleration

Empirical studies demonstrate that RRT* can require hundreds of thousands to millions of nodes to approximate an optimal path, especially in high-cost or narrow regions (Qureshi et al., 2017, Qureshi et al., 2017, Tahir et al., 2018). Uniform sampling, while unbiased, leads to slow refinement near the optimal corridor, resulting in high memory and execution time. These drawbacks are particularly pronounced in online or real-time applications, where rapid convergence is critical (Qureshi et al., 2017, T et al., 2024).

3. Advanced Variants: Potential-Guided and Bidirectional RRT*

To mitigate slow convergence, extensions of RRT* introduce domain-specific heuristics into the sampling, connection, or rewiring phases.

3.1 Potential-Function Based RRT* (P-RRT*)

P-RRT* integrates an artificial potential field (APF) heuristic to bias sampling toward the goal:

• Attractive potential: $U_\mathrm{att}(x) = \|x - x_\mathrm{goal}\|^2$
• Randomized Gradient Descent: For each sample $x_\mathrm{rand}$,

$$x_\mathrm{prand} \leftarrow x_\mathrm{rand}; \quad \text{for } k \text{ steps: } x_\mathrm{prand} \leftarrow x_\mathrm{prand} + \lambda \frac{-\nabla U_\mathrm{att}}{\|\nabla U_\mathrm{att}\|}$$

stopped early if the point nears an obstacle, improving sample density near low-cost corridors but retaining global completeness.

This mechanism accelerates convergence, reducing required samples and memory usage by one to two orders of magnitude in both 2D and 3D simulated scenarios (Qureshi et al., 2017). In 2D maze environments, average required samples dropped from $\sim4.0 \times 10^6$ (RRT*) to $\sim1.5 \times 10^5$ (P-RRT*), with similar reductions in runtime and memory.

3.2 Bidirectional and Intelligent Bidirectional RRT*

• B-RRT*: Grows two trees—one from the initial state, one from the goal—attempting connection when either tree expands. This halves the constant in the convergence rate and expedites path discovery in cluttered spaces (Qureshi et al., 2017, Tahir et al., 2018).
• IB-RRT*: Uses a heuristic to determine which tree to expand by evaluating which root is closer in cumulative cost to a candidate sample. The connection is only attempted if both trees have nodes within the standard near-radius.

Bidirectional approaches empirically yield one to two orders of magnitude faster convergence to the optimal path in challenging environments compared to single-tree RRT* (Qureshi et al., 2017). For example, in 3D mazes, IB-RRT* required $\sim1.7 \times 10^5$ samples versus $2.2 \times 10^6$ for RRT*, reflecting a significant speedup.

3.3 Potentially Guided Bidirectional RRT*

By integrating potential-field bias within each tree of a bidirectional scheme (PIB-RRT*, PB-RRT*), the convergence rate is further improved. In 50-run experiments, PIB-RRT* achieved median times to within $0.1\%$ of optimum in $\sim1.8$ s versus $8.2$ s for vanilla RRT* (Tahir et al., 2018).

4. Convergence, Completeness, and Complexity

All RRT* variants described here retain theoretical guarantees of probabilistic completeness and asymptotic optimality provided they maintain:

• Sampling with non-zero density in $X_\mathrm{free}$
• A connection radius $r_n = O((\log n / n)^{1/d})$ that ensures connectivity
• Rewiring steps that strictly lower cost

The per-iteration computational complexity is $O(\log n)$ for neighbor searches (e.g., with a $k$-d tree) and $O(n)$ for total memory (Qureshi et al., 2017, Tahir et al., 2018, Qureshi et al., 2017).

Bidirectional and potential-guided methods reduce the constant in the convergence-rate bound. For standard RRT*, expected suboptimality decreases as $O((\log N / N)^{1/d})$. Both potential-biased and bidirectional extensions halve this constant or better in cluttered/narrow spaces (Tahir et al., 2018).

5. Quantitative Performance and Experimental Results

The following table synthesizes comparative results for representative algorithms across multiple environments (Qureshi et al., 2017, Qureshi et al., 2017, Tahir et al., 2018):

Scenario	RRT* Avg Samples / Time / Failures	Advanced Variant	Avg Samples / Time / Failures	Path Cost
2D Maze A	4.0M / 260 s / 7	P-RRT*	150k / 29 s / 0	≈163
3D Barriers	1.96M / 128 s / 11	IB-RRT*	204k / 47.8 s / 0	≈82
2D Cluttered	Fails (>5M nodes)	P-RRT*	3.0k / 0.56 s / 0	≈39
All (cluttered)	high (1–2 orders higher runtime)	PIB-RRT*	5,500 nodes @10s / 1.8 s	$c^*$

Bidirectional and potential-biased methods consistently required fewer nodes, less memory, and reached optimality faster.

6. Extensions for Dynamic and Non-Holonomic Systems

• Dynamic Environments: Bidirectional and potential-guided algorithms (e.g., Bi-AM-RRT* (Zhang et al., 2023)) with enhanced rewiring schemes support dynamic obstacle avoidance in real time by locally rewiring existing branches, without full replanning.
• Non-Holonomic Constraints: P-RRT* may use kinodynamic steering (e.g., Runge–Kutta integration for differential-drive robots). Completeness and optimality still hold, and iteration reductions of $\geq 50\times$ have been demonstrated in repeated Pioneer 3-DX experiments (Qureshi et al., 2017).

7. Applications, Limitations, and Future Directions

Applications of RRT* and its variants span autonomous robotics (UAV collision avoidance, parking, autonomous driving), where memory efficiency and convergence rate are critical (Qureshi et al., 2017, Killian et al., 2021, T et al., 2024).

Limitations include:

• Sensitivity to sampling and connection hyperparameters, especially in bidirectional schemes (e.g., connection radius $\sigma$ in Bi-AM-RRT* must be finely tuned) (Zhang et al., 2023).
• Potential-guided schemes may incur traps near local minima without synergistic global exploration.
• Discarding tree branches after bidirectional connection can slow reoptimization for new goals or dynamically changing environments (Zhang et al., 2023).

Future research focuses on hybridizing bidirectional, potential-guided, and metric-assisting sampling, and retaining partial subtrees for continual replanning in non-stationary domains.

---

References:

• "Potential Functions based Sampling Heuristic For Optimal Path Planning" (Qureshi et al., 2017)
• "Intelligent bidirectional rapidly-exploring random trees for optimal motion planning in complex cluttered environments" (Qureshi et al., 2017)
• "Potentially Guided Bidirectionalized RRT* for Fast Optimal Path Planning in Cluttered Environments" (Tahir et al., 2018)
• "Bi-AM-RRT*: A Fast and Efficient Sampling-Based Motion Planning Algorithm in Dynamic Environments" (Zhang et al., 2023)
• "Utilizing the RRT*-Algorithm for Collision Avoidance in UAV Photogrammetry Missions" (Killian et al., 2021)
• "Modified RRT* for Path Planning in Autonomous Driving" (T et al., 2024)