Direction Informed Trees (DIT*)

• Direction Informed Trees (DIT*) is a sampling-based optimal motion planner that extends BIT*, AIT*, and EIT* by incorporating explicit directional estimation for candidate edges.
• It uses a robust directional filter and cost heuristic to prune unhelpful neighbors before expensive collision checks, enhancing search efficiency.
• The method improves initial solution time and path quality while preserving probabilistic completeness and almost-sure asymptotic optimality in high-dimensional planning.

Direction Informed Trees (DIT) is a single-query, sampling-based optimal motion planner that extends the informed-tree lineage of BIT, AIT, and especially EIT by explicitly estimating search direction for each candidate edge. In DIT*, edges are treated as generalized vectors in configuration space, compared through similarity indexes, filtered before expensive validation, and assigned a direction cost heuristic that biases exploration toward goal-consistent directions. The method is presented as a way to complement cost and effort heuristics with an explicit notion of directional usefulness, while retaining the probabilistic completeness and almost-sure asymptotic optimality associated with informed random-geometric-graph planners (Zhang et al., 26 Aug 2025).

1. Problem setting and algorithmic lineage

DIT* is formulated in the standard continuous optimal path-planning setting over a state space $X \subseteq \mathbb{R}^n$, with obstacle region $X_{\mathrm{obs}} \subset X$, free space

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$

start state $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$, and goal set $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$. A feasible path is a continuous map $\sigma:[0,1]\to X$, and the objective is

$\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$

This is the same asymptotically optimal planning regime used by RRT*, BIT*, AIT*, and EIT* (Zhang et al., 26 Aug 2025).

The planner is positioned as an extension of the informed-tree family rather than as a new graph model. The paper states that DIT* keeps the same choose-parent and rewire strategies as EIT*, and preserves the asymmetric bidirectional organization of informed planners built on implicit random geometric graphs. The claimed gap is heuristic: BIT* uses admissible cost ordering, AIT* adds adaptive heuristic estimation, EIT* adds effort-aware reverse guidance, and DIT* argues that these mechanisms still do not explicitly evaluate whether a candidate edge points in a promising search direction (Zhang et al., 26 Aug 2025).

The central motivation is operational. In single-query optimal planning, nearest-neighbor processing and collision checking are major costs, with collision checking emphasized as especially expensive. DIT* therefore attempts to eliminate directionally unhelpful neighbors before exact edge validation and to prefer locally goal-consistent edges when multiple candidates have similar effort or cost profiles. This makes DIT* less a change to the sampling model than a change to neighbor selection and edge prioritization (Zhang et al., 26 Aug 2025).

2. Directional representation and heuristic machinery

The defining move in DIT* is to represent an edge as a generalized vector in configuration space. For an edge between source state $\mathbf{x}_s$ and target state $\mathbf{x}_t$, the paper writes

$$\mathcal{V}ec := V_{\mathcal{F}}(\mathbf{x}_s) - V_{\mathcal{F}}(\mathbf{x}_t), \qquad \mathcal{V}ec \in \mathbb{R}^{n\times 1},$$

with norm and normalization

$X_{\mathrm{obs}} \subset X$0

The notation is awkwardly typeset in the paper, but the intended meaning is the displacement associated with an edge (Zhang et al., 26 Aug 2025).

At a forward-search state $X_{\mathrm{obs}} \subset X$1, DIT* uses three vectors. The target vector

$X_{\mathrm{obs}} \subset X$2

encodes the previous motion. The check vector

$X_{\mathrm{obs}} \subset X$3

encodes the candidate motion. The goal vector

$X_{\mathrm{obs}} \subset X$4

encodes the current goalward direction. The geometric intent is that a candidate neighbor is favorable if its outgoing direction remains consistent with, or improves upon, the current trajectory’s alignment to the goal (Zhang et al., 26 Aug 2025).

Directional comparison is performed with a weighted cosine-like similarity index,

$X_{\mathrm{obs}} \subset X$5

with $X_{\mathrm{obs}} \subset X$6 and prior weight parameters $X_{\mathrm{obs}} \subset X$7. The filter computes

$X_{\mathrm{obs}} \subset X$8

The paper’s prose describes this as retaining neighbors that are more goal-oriented and filtering those that are less favorable (Zhang et al., 26 Aug 2025).

The second directional mechanism is the direction cost heuristic. The paper gives

$X_{\mathrm{obs}} \subset X$9

and Algorithm 2 instantiates this with

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$0

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$1

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$2

If either $$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$3 or $$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$4, the edge is assigned zero direction cost. The paper distinguishes admissible quantities, denoted with hats, from potentially inadmissible ones, denoted with bars; in practice the directional penalty used in queue ordering is the estimated $$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$5 (Zhang et al., 26 Aug 2025).

The paper also introduces directional path terms

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$6

together with admissible counterparts. This establishes direction as a third heuristic axis, alongside path cost and validation effort (Zhang et al., 26 Aug 2025).

3. Search process, filtering, and queue organization

DIT* is operationally an EIT*-style planner with directional augmentation. It maintains a forward tree

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$7

and a reverse tree

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$8

together with forward and reverse edge queues. Sampling is uniform, the implementation uses batch methods, and in the reported experiments the batch planners use 100 states per batch (Zhang et al., 26 Aug 2025).

The reverse-search queue is ordered lexicographically by cost, effort, and direction. The paper writes the reverse key as

$$X_{\mathrm{free}} = cl(X \setminus X_{\mathrm{obs}}),$$9

In the intended use, the third key becomes decisive when effort differences are sufficiently small. The paper states the threshold condition as

$\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$0

with experimental value $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$1 (Zhang et al., 26 Aug 2025).

The forward-selection rule mirrors this. Among forward-queue edges with minimum estimated effort, DIT* chooses the edge with the smallest direction cost, equivalently the smallest directional path estimate. The paper writes

$\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$2

and its inadmissible counterpart

$\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$3

Algorithm 3 then selects the edge minimizing $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$4, or equivalently $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$5, when effort is comparable (Zhang et al., 26 Aug 2025).

The distinctive procedural step is early directional filtering. For each candidate neighbor of the current source, DIT* constructs $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$6, $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$7, and $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$8, computes $\mathbf{x}_{\mathrm{start}} \in X_{\mathrm{free}}$9 and $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$0, and discards neighbors that fail the directional test before exact edge validation. The intended benefit is fewer collision checks and a smaller edge queue. The paper includes a safety condition for near-zero vectors: $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$1 After surviving the filter, edges are processed using the same choose-parent and rewiring logic as EIT* (Zhang et al., 26 Aug 2025).

The paper does not present one monolithic planner pseudocode. Instead, DIT* is described as an EIT*-like forward/reverse informed planner in which reverse search continues to provide admissible and inadmissible cost and effort estimates, while the added directional terms alter neighbor selection and tie-breaking. This suggests that DIT* should be understood less as a new tree topology and more as a direction-aware scheduling policy over the EIT* search structure (Zhang et al., 26 Aug 2025).

4. Guarantees and analytical status

The paper claims that DIT* preserves the standard guarantees of uniform-sampling optimal planners. For probabilistic completeness it states

$X_{\mathrm{goal}} \subset X_{\mathrm{free}}$2

For asymptotic optimality it uses the standard rewiring-radius condition

$X_{\mathrm{goal}} \subset X_{\mathrm{free}}$3

where $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$4 is the number of sampled states, $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$5 is a tuning parameter, $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$6 is dimension, $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$7 is the Lebesgue measure of the informed set, and $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$8 is the volume of the unit ball in $X_{\mathrm{goal}} \subset X_{\mathrm{free}}$9. Under this condition, the paper states

$\sigma:[0,1]\to X$0

(Zhang et al., 26 Aug 2025).

The theoretical argument is explicitly inheritance-based. DIT* uses uniform sampling and the same choose-parent and rewire strategies as EIT*, so the direction filter and direction cost are presented as search accelerators that do not alter the underlying asymptotic graph-density conditions. The paper does not provide a new convergence-rate theorem for the directional heuristic, nor a separate proof that the directional cost is admissible or consistent. On the contrary, it explicitly permits $\sigma:[0,1]\to X$1 to be inadmissible (Zhang et al., 26 Aug 2025).

This analytical posture is important. DIT* does not claim that the direction model changes the asymptotic order of convergence in the formal RGG sense. Instead, it claims that directional information improves practical search efficiency while the standard guarantees remain available because the underlying sampling and rewiring machinery is unchanged. A plausible implication is that DIT* should be read as a heuristic-strengthening variant of EIT*, not as a new theoretical class of sampling-based planner.

5. Experimental evidence and reported performance

The implementation was integrated into OMPL and evaluated against official OMPL implementations of RRT-Connect, Informed RRT*, BIT*, AIT*, ABIT*, and EIT*. Simulation experiments were run in

$\sigma:[0,1]\to X$2

using two synthetic environments: Random Rectangles (RR) and Narrow Passage (NP). Each instance was run 100 times per planner. The reported settings include RGG constant $\sigma:[0,1]\to X$3, rewire factor $\sigma:[0,1]\to X$4, RRT-based goal bias $\sigma:[0,1]\to X$5, maximum edge lengths $\sigma:[0,1]\to X$6, $\sigma:[0,1]\to X$7, and $\sigma:[0,1]\to X$8 for $\sigma:[0,1]\to X$9, $\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$0, and $\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$1, and 100 samples per batch for batch planners. DIT* uses

$\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$2

in the reported experiments (Zhang et al., 26 Aug 2025).

The primary reported metrics are median initial solution time $\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$3, median initial solution cost $\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$4, median final solution cost $\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$5, and success rate. Across the synthetic benchmarks, the dominant pattern is that DIT* improves both initial-solution time and path quality relative to EIT*, with gains increasing in higher dimension and under stronger directional constraints such as narrow passages (Zhang et al., 26 Aug 2025).

Benchmark	EIT* $\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$6	DIT* $\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$7
RR-$\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$8	0.0660 / 2.1817 / 1.2943	0.0585 / 1.8194 / 1.2040
RR-$\begin{split} \sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \;\middle|\; \sigma(0) = \mathbf{x}_{\text{start}},\ \sigma(1) \in X_{\text{goal}}, \right. \ &\qquad\qquad \left. \forall t \in [0, 1],\ \sigma(t) \in X_{\text{free}} \right\}. \end{split}$9	0.1221 / 3.9503 / 2.6709	0.0958 / 3.0148 / 2.2551
RR-$\mathbf{x}_s$0	0.2047 / 5.7269 / 3.9990	0.1514 / 4.0179 / 3.1000
NP-$\mathbf{x}_s$1	0.0165 / 1.1758 / 0.5373	0.0110 / 1.1095 / 0.4944
NP-$\mathbf{x}_s$2	0.0346 / 2.9125 / 2.1870	0.0225 / 2.3104 / 1.6899
NP-$\mathbf{x}_s$3	0.0712 / 4.7750 / 3.7961	0.0391 / 3.4845 / 2.9378

The paper highlights median initial-time improvements of $\mathbf{x}_s$4, $\mathbf{x}_s$5, and $\mathbf{x}_s$6 on RR-$\mathbf{x}_s$7, RR-$\mathbf{x}_s$8, and RR-$\mathbf{x}_s$9, and $\mathbf{x}_t$0, $\mathbf{x}_t$1, and $\mathbf{x}_t$2 on NP-$\mathbf{x}_t$3, NP-$\mathbf{x}_t$4, and NP-$\mathbf{x}_t$5. It also notes that in RR-$\mathbf{x}_t$6, DIT*’s $\mathbf{x}_t$7 is $\mathbf{x}_t$8 shorter than EIT*’s (Zhang et al., 26 Aug 2025).

The paper also reports two real-world manipulation tasks.

Task	EIT*	DIT*
Kitchen model, 7-DoF	93.33% success, cost 22.71	100% success, cost 16.84
Dual-arm cable routing, 14-DoF	90.00% success, cost 21.85	96.67% success, cost 15.48

In the kitchen model, a robotic arm grasps and relocates a fork in a cluttered environment with a 3.0 s time limit over 30 trials. In the dual-arm cable-routing task, a 14-DoF system plans around fixture-holder obstacles with a 5.0 s time limit over 30 runs. In both cases DIT* reports the highest success rate and the shortest median path length among the listed informed-tree baselines (Zhang et al., 26 Aug 2025).

The experimental picture is nonetheless incomplete. The paper does not include a full ablation study isolating the direction filter from the direction-cost heuristic, and it does not systematically vary $\mathbf{x}_t$9, $$\mathcal{V}ec := V_{\mathcal{F}}(\mathbf{x}_s) - V_{\mathcal{F}}(\mathbf{x}_t), \qquad \mathcal{V}ec \in \mathbb{R}^{n\times 1},$$0, $$\mathcal{V}ec := V_{\mathcal{F}}(\mathbf{x}_s) - V_{\mathcal{F}}(\mathbf{x}_t), \qquad \mathcal{V}ec \in \mathbb{R}^{n\times 1},$$1, $$\mathcal{V}ec := V_{\mathcal{F}}(\mathbf{x}_s) - V_{\mathcal{F}}(\mathbf{x}_t), \qquad \mathcal{V}ec \in \mathbb{R}^{n\times 1},$$2, or $$\mathcal{V}ec := V_{\mathcal{F}}(\mathbf{x}_s) - V_{\mathcal{F}}(\mathbf{x}_t), \qquad \mathcal{V}ec \in \mathbb{R}^{n\times 1},$$3. The reported evidence therefore shows that the combined method is effective on the tested benchmarks, but not which directional component is principally responsible for the gain (Zhang et al., 26 Aug 2025).

6. Related planners, limitations, and terminological scope

Within the informed-tree lineage, BIT* establishes batchwise heuristic search over an implicit random geometric graph (Gammell et al., 2017), and AIT* shows how an asymmetric reverse search can estimate a stronger problem-specific heuristic online (Strub et al., 2020). DIT* is presented as EIT* plus explicit direction estimation: it leaves the parent-selection and rewiring machinery intact, but adds a direction filter for early neighbor pruning and a direction cost heuristic for edge prioritization (Zhang et al., 26 Aug 2025).

Several later planners explore adjacent design axes rather than explicit directionality. FIT* changes the batch-size policy over time rather than the edge direction model (Zhang et al., 2023). FCIT* removes nearest-neighbor search entirely by searching a fully connected implicit graph with informed ordering (Wilson et al., 2024). GIT* replaces hand-designed reverse heuristics with reinforced genetic-programming expressions that incorporate obstacle-derived and search-history-derived signals (Zhang et al., 28 Aug 2025). This suggests that “informed trees” has evolved into a family that varies along heuristic design, connectivity policy, and batch scheduling, with DIT* occupying the specific niche of local directional filtering.

DIT* also has evident limitations. The paper itself notes notation and formulation issues: the direction-filter pseudocode appears to contain a sign inconsistency, the admissible and inadmissible direction symbols are used somewhat loosely, and the printed weighted cosine similarity is unconventional. More substantively, the planner is most likely to help when local goalward alignment is a useful predictor of good edges; the paper explicitly notes that it may be less beneficial when obstacle geometry requires substantial detours that temporarily move away from the goal, when candidate-neighbor sets are already small, or when directional similarity is poorly correlated with feasibility. The paper does not analyze such adversarial cases in depth (Zhang et al., 26 Aug 2025).

The term itself is also not universal across the literature. In integrated task and motion planning, TMIT* uses asymmetric forward/reverse informed search in a hybrid symbolic-geometric space, but it is not a DIT* variant (Thomason et al., 2022). In active gas-distribution mapping, XIT grows a tree using a UCB information field and frontier goals, again without using the DIT* directional-vector formulation (Fazliu et al., 14 Feb 2026). Outside motion planning altogether, high-dimensional spatial data structures can be “direction informed” in the sense that split normals are derived from local variance structure, as in Approximate Principal Direction Trees (McCartin-Lim et al., 2012). Conversely, the acronym DIT is unrelatedly used for the Double Interdiction Problem on Trees, a rooted-tree optimization problem with no connection to informed sampling-based planning (Li et al., 2024).

In that broader context, DIT* is most precisely understood as a direction-aware extension of the BIT*/AIT*/EIT* tradition: it inserts explicit local orientation reasoning into the queue-based, asymptotically optimal search over sampled graphs. Its distinctive claim is not that search should merely be informed by cost, effort, or application-specific rewards, but that the geometric direction of each candidate edge is itself a reusable planning signal.