Bi-Directional Space-Time RRT Planning

• Bi-directional space-time RRT planning is a framework that concurrently grows two search trees in a joint configuration-time space to effectively manage dynamic obstacles and kinodynamic constraints.
• It employs strategies like informed sampling, hybrid dynamics handling, motion primitives, and safe interval verification to rapidly explore feasible trajectories.
• Empirical evaluations reveal that these methods offer improved probabilistic completeness, reduced computation times, and enhanced solution quality compared to unidirectional or static-environment planners.

Bi-directional space-time Rapidly-exploring Random Tree (RRT) planning encompasses a set of algorithms that address motion planning through configuration space augmented by a time dimension, explicitly handling dynamic obstacles, temporal constraints, and system dynamics. This paradigm generalizes single-direction RRT approaches by coordinating the growth of two search trees—one forward from the start and one backward from the goal—within the joint configuration-time manifold. Algorithmic representatives include ST-RRT* for time-optimal planning with velocity constraints, HyRRT-Connect for hybrid (continuous-discrete) dynamics, iDb-RRT for kinodynamic systems via motion primitives and trajectory repair, and SI-RRT for randomized planning with safe intervals in the presence of dynamic obstacles. These frameworks yield improved completeness, solution quality, and practical efficiency over unidirectional or static-environment counterparts.

1. Problem Formulation and State Space Construction

The foundational model introduces an augmented state space $X = Q \times T$, where $Q$ is the robot's configuration space (e.g., $\mathbb{R}^n$ or a more general manifold), and $T$ is a time interval, typically $[0, \infty)$ or $[0, t_{\max}]$ for a finite horizon. At each time $t \in T$, the set of feasible configurations evolves according to dynamic obstacle trajectories, and system-specific constraints (kinodynamic bounds, hybrid transitions) further restrict feasibility.

In ST-RRT*, state validity requires that $x = (q, t)$ is collision-free at $t$, and transitions must obey velocity constraints such that $|q_2^i - q_1^i| \le v^i_{\max} (t_2 - t_1)$ for axis $i$, $t_2 > t_1$ (Grothe et al., 2022). In HyRRT-Connect, the state is a tuple $(x, t, j)$ with $j$ denoting the discrete mode of a hybrid system and trajectories comprise piecewise-continuous flows interleaved with discrete jumps, defined on hybrid time domains $\mathcal{T} \subset \mathbb{R}_+ \times \mathbb{N}$ (Wang et al., 14 Apr 2025, Wang et al., 2024).

Safe-Interval RRT planners such as SI-RRT instead maintain for each $q \in Q$ a set of maximal intervals $S(q)$ over which $q$ is collision-free with respect to dynamic obstacles—a temporally chunked abstraction replacing explicit time-stamping of each node (Kerimov et al., 2024).

2. Bidirectional Expansion and Connection Strategies

Bidirectional expansion underpins efficiency and probabilistic completeness in these planners. Two trees are grown—one forward in time from the initial state or set, the other backward from the goal region (with time reversed if the dynamics admit inversion):

• ST-RRT*: Alternates tree roles each iteration. In each tree, a random state in $Q \times T$ is sampled via informed strategies. The forward tree extends from $x_{\text{start}}$; the backward tree comprises a batch of sampled goal states from $Q_{\text{goal}} \times T_{\text{goal}}$, progressively expanded until a solution is found (Grothe et al., 2022).
• HyRRT-Connect: Propagates both trees under hybrid system dynamics—forward tree under original system, backward under time-reversed system. Overlap is detected either in continuous flow (within $\delta$ in state) or via perfect jump matching, and a candidate solution is assembled by reversal and concatenation of the respective trajectories (Wang et al., 14 Apr 2025, Wang et al., 2024).
• iDb-RRT: Allows $\delta$-discontinuities when connecting motion primitives and grows both trees, bridging once stop conditions are met within a bounded state-time distance, postponing exact feasibility to a trajectory optimization phase (Ortiz-Haro et al., 2024).
• SI-RRT: Grows trees in configuration space only, assigning arrival intervals post-facto via fast safe interval computation. Trees connect via matching configurations at compatible time intervals and at compatible parent node arrival times, followed by a unification of path segments (Kerimov et al., 2024).

Below is an organizational summary:

Algorithm	State Representation	Connection/Unification
ST-RRT*	$(q, t)$	Extend/Connect with pruning; swap trees
HyRRT-Connect	$(x, t, j)$	Overlap (flow/jump), concatenation+reversal
iDb-RRT	$(x, t)$ with primitives	$\delta$-discontinuity, optimization repair
SI-RRT	$q$ with intervals	Interval-based, configuration meet

3. Sampling and Search Heuristics

Randomized sampling is coupled with problem-specific heuristics to accelerate convergence:

• Informed Sampling (ST-RRT*): Samples are drawn from the intersection of two reachability cones defined by velocity limits; only $(q, t)$ pairs for which $t \in [t_{\min}(q), t_{\max}(q)]$ are attempted, substantially pruning infeasible regions (Grothe et al., 2022).
• Goal Region Expansion (ST-RRT*): The temporal component of $X_{\text{goal}}$ is grown in geometrically increasing batches (controlled via rangeFactor $\alpha$ and initialBatchSize), ensuring unknown optimal arrival times $T^*$ are eventually included and preventing pathological oversampling (Grothe et al., 2022).
• Hybrid Time Handling (HyRRT-Connect): Samples may be biased toward flow regions or toward jump surfaces, allowing rapid exploration of hybrid dynamics and facilitating overlap conditions (Wang et al., 14 Apr 2025).
• Motion Primitive Libraries (iDb-RRT): Instead of arbitrary controls, trees expand via libraries of precomputed, optimal local trajectories, enabling fine-grained control of state-time transitions and simplifying dynamic feasibility (Ortiz-Haro et al., 2024).
• Safe Intervals (SI-RRT): Each configuration sample is efficiently labeled with its maximal time intervals of safety, removing the need to sample explicit $t$ values and allowing rapid interval-based attachment and connection strategies (Kerimov et al., 2024).

4. Solution Construction, Optimality, and Theoretical Guarantees

Once overlap or connection is detected, the algorithms construct a feasible, often near-optimal, solution trajectory:

• ST-RRT*: On successful connection, the solution cost provides an upper bound for further search; trees are pruned outside feasible time cones. Only the goal trees are rewired to minimize arrival time, ensuring asymptotic optimality as the number of samples grows (Grothe et al., 2022).
• HyRRT-Connect: Paths from both trees are combined—if a small $\delta$ gap exists, a forward-in-time resimulation with the reversed backward input is performed, ensuring vanishing end-state error as $\delta \to 0$. Connections by exact jump are used wherever possible to avoid discontinuity (Wang et al., 14 Apr 2025, Wang et al., 2024).
• iDb-RRT: A discontinuous path is optimized post-hoc via a trajectory optimizer (e.g., DD-P), enforcing true dynamics, collision, and other constraints. If repair fails, the allowed discontinuity is decreased and the process repeats (Ortiz-Haro et al., 2024).
• SI-RRT: Path assembly involves splicing two tree segments matched at configuration and compatible safe intervals. Time flows forward along the stitched trajectory, and the resulting path is guaranteed feasible given proper assignment of parent intervals and arrival times (Kerimov et al., 2024).

Analytically, probabilistic completeness is shown for all frameworks under standard regularity conditions—feasible regions are open, the sampling domain is sufficiently rich, and controls are sufficiently dense. Asymptotic optimality is only guaranteed for approaches where explicit rewiring or trajectory repair drives cost to the infimum, e.g., ST-RRT*; SI-RRT in its basic form is probabilistically complete but not necessarily optimal, though an asymptotically-optimal variant with internal rewiring is suggested (Grothe et al., 2022, Kerimov et al., 2024).

5. Computational Performance and Empirical Evaluation

Across diverse benchmarks, bi-directional space-time RRT planners outperform prior single-tree or static-environment methods, particularly in temporal bottleneck scenarios and high-dimensional spaces:

• ST-RRT: Consistently obtained 100% success across all benchmarks. Median times to first solution were 2–10$\times$ faster than space-time RRT, and final solution costs (arrival times) were lower, especially in cluttered or narrow scenarios. RRT* shows slower convergence, RRT-Connect does not refine once a solution is found. All algorithms executed on OMPL, under 30–100 seconds per run (Grothe et al., 2022).
• HyRRT-Connect: Demonstrated order-of-magnitude reductions in vertex count and wall-time compared to single-tree or unidirectional methods. In the actuated bouncing ball benchmark, 0.27 s (HyRRT-Connect) vs. 0.76 s (Bi-HyRRT) vs. 3.93 s (HyRRT) (Wang et al., 14 Apr 2025, Wang et al., 2024). On the compass-gait biped, 19.8 s versus 71.5 s with comparable success rates.
• iDb-RRT (bidirectional): Achieved first solutions up to 10$\times$ faster than prior kinodynamic planners such as OMPL/Kino-RRT. Typical solution rates exceeded 90%, with average cost (time) within 1.2$\times$ of discrete-search baselines across 30 real-world tasks (Ortiz-Haro et al., 2024).
• SI-RRT: Scaled to scenarios with hundreds of moving obstacles (e.g., 220 obstacles at 77% success, compared to 42% for ST-RRT*, and 6% for DRGBT). In all tests, SI-RRT achieved dramatically lower runtimes (e.g., at 100 obstacles: 1.2 s vs. 12.5 s for ST-RRT*) and lower solution cost for arrival time (e.g., 4.8 s vs. 6.5 s)—an order-of-magnitude improvement in both metrics (Kerimov et al., 2024).

6. Limitations and Research Directions

Existing approaches share certain common limitations:

• Parameter tuning (e.g., rangeFactor $\alpha$ in ST-RRT*, $\delta$ in iDb-RRT, safe interval resolution in SI-RRT) is nontrivial and can impact both efficiency and solution quality (Grothe et al., 2022, Ortiz-Haro et al., 2024, Kerimov et al., 2024).
• ST-RRT* and SI-RRT rely on prior knowledge of obstacle trajectories and assume velocity-bounded kinodynamic models; higher-order constraints (acceleration, force limits) and nonholonomic constraints are not yet systematically addressed (Grothe et al., 2022, Kerimov et al., 2024).
• No adaptation to online changes in the dynamic environment is implemented, although SI-RRT suggests leveraging existing tree and interval structure for anytime replanning should obstacles deviate from predicted motion (Kerimov et al., 2024).
• Bi-directional connection in hybrid systems requires careful handling of discrete jump alignment; resolving flow discontinuities requires additional simulation and may yield small approximation errors, though these can be made arbitrarily small (Wang et al., 14 Apr 2025, Wang et al., 2024).

Promising future research directions include:

• Incorporating trajectory optimization or rewiring within safe intervals for combined completeness and asymptotic optimality (SI-RRT extension) (Kerimov et al., 2024).
• Handling richer kinodynamic constraints via expanded state spaces or integration with direct trajectory optimization frameworks (Grothe et al., 2022).
• Adaptive batch sizing and interval refinement to autonomously tune planner parameters (Grothe et al., 2022, Kerimov et al., 2024).
• Extension to multi-agent or coupled systems via interval-graph coordination or decentralized bi-directional search (Kerimov et al., 2024).

7. Comparative Summary

The emergence of bi-directional space-time RRT-style planners marks a substantial advance in motion planning under temporal and kinodynamic constraints. The following table summarizes algorithmic focus and domain features.

Planner	Dynamics Model	Time Handling	Key Domain Strength
ST-RRT*	Velocity-bounded	Continuous, unbounded	Asymptotic optimality, no a priori arrival time required (Grothe et al., 2022)
HyRRT-Connect	Hybrid (flows+jumps)	Hybrid time domain	Fast connection in hybrid systems, continuity repair (Wang et al., 14 Apr 2025, Wang et al., 2024)
iDb-RRT	General kinodynamic	State-time	Motion primitives, fast kinodynamic planning, trajectory repair (Ortiz-Haro et al., 2024)
SI-RRT	High DOF+dynamic obs.	Safe intervals	Efficient manipulation with hundreds of obstacles, order-of-magnitude speedup (Kerimov et al., 2024)

Each method leverages a distinct representation of time, search heuristics, and solution assembly, but all share the bidirectional paradigm for efficient, probabilistically complete exploration in space-time. As new domains require planning under even more demanding kinodynamic and temporal uncertainty structures, these frameworks provide both foundational and practical tools for scalable, high-quality decision making in dynamical environments.