---
title: Multi-Informed Trees (MIT*) Planning
url: https://www.emergentmind.com/topics/multi-informed-trees-mit
type: topic
---

# Multi-Informed Trees (MIT*) Planning

Multi-Informed Trees (MIT*) is a single-query, sampling-based motion planner for continuous, high-dimensional configuration spaces that addresses a specific limitation of prior informed planners: before a first feasible path is found, standard informed-set methods cannot condition sampling on a finite incumbent cost and therefore revert to global exploration. MIT* introduces an estimated informed set before the initial solution, derived from a prior admissible solution cost obtained from lazy reverse-search failure and then expanded by a reliability-dependent factor; after a feasible path exists, it transitions to the standard informed set defined by the current best solution cost. The planner combines a forward tree of valid states and valid edges, a lazy reverse tree whose edges may traverse invalid regions, an adaptive sampler, and length-related adaptive sparse collision checks, while claiming probabilistic completeness and asymptotic optimality [2508.21549].

## 1. Definition and planning setting

MIT* is formulated for optimal path planning in a state space \(X \subseteq \mathbb{R}^n\) with obstacle set \(X_{\textnormal{obs}}\), collision-free space \(X_{\textnormal{free}}\), a start state \(\mathbf{x}_{\textnormal{start}}\), and a goal region \(X_{\textnormal{goal}}\). A path is a continuous map \(\sigma : [0,1] \mapsto X\), and the optimal path is written as
\[
\begin{split}
\sigma^* &= \arg \min_{\sigma \in \Sigma} \left\{ c(\sigma) \,\middle|\, \sigma(0) = \mathbf{x}_{\textnormal{start}}, \sigma(1) \in \mathbf{x}_{\textnormal{goal}}, \right. \\
&\qquad\qquad \left. \forall t \in [0, 1], \sigma(t) \in X_{\textnormal{free}} \right\}.
\end{split}
\]
In the reported experiments, the objective is path length [2508.21549].

The planner is motivated by the behavior of informed methods that define an informed subset only after a feasible solution exists. With admissible estimates
\[
\hat{g}(\mathbf{x}) := \hat{c}(\mathbf{x}_{\textnormal{start}}, \mathbf{x}), \qquad
\hat{h}(\mathbf{x}) := \min_{\mathbf{x}_{\textnormal{goal}} \in X_{\textnormal{goal}}}\hat{c}(\mathbf{x}, \mathbf{x}_{\textnormal{goal}}),
\]
and
\[
\hat{f}(\mathbf{x}) := \hat{g}(\mathbf{x}) + \hat{h}(\mathbf{x}),
\]
the standard informed set is
\[
X_{\hat{f}} := \{\mathbf{x} \in X \mid \hat{f}(\mathbf{x}) < c_\textnormal{curr}\},
\]
where \(c_\textnormal{curr}\) is the current solution cost. This mechanism is effective after a feasible solution exists, but before that stage standard informed planners revert to sampling the whole space. MIT* is designed to recover focused search in that pre-solution phase by constructing an estimated informed set from prior admissible costs [2508.21549].

The “multi-informed” designation refers to the use of multiple informed subsets or guidance mechanisms over time. MIT* maintains a forward tree \(\mathcal{T_F} = (V_\mathcal{F}, E_\mathcal{F})\), a reverse tree \(\mathcal{T_R} = (V_\mathcal{R}, E_\mathcal{R})\), an estimated informed set \(X_{\tilde{\sigma}}\) before any feasible solution exists, and the standard informed set \(X_{\hat f}\) after a valid solution has been found. This temporal transition between pre-solution and post-solution informed subsets is the central structural distinction of the method [2508.21549].

## 2. Estimated informed sets and the meaning of “multi-informed”

MIT* defines the underlying cost structure through
\[
g(x) := \min_{\sigma \in \Sigma} \{c(\sigma) \mid \sigma(0) = x_{\textnormal{start}}, \sigma(1) = x\},
\]
\[
h(x) := \min_{\sigma \in \Sigma} \{c(\sigma) \mid \sigma(0) = x, \sigma(1) \in X_{\textnormal{goal}}\},
\]
and
\[
f(x) := g(x) + h(x).
\]
The central innovation is a prior admissible cost available before the first valid path is known. When a forward-search attempt fails full collision checking after connecting the forward tree to the reverse tree, MIT* constructs an admissible initial solution cost
\[
s := g_{\mathcal{F}(\mathbf{x}_s) + \hat{c}(\mathbf{x}_s,\mathbf{x}_t) + \hat{h}(\mathbf{x}_t),
\]
where \(g_{\mathcal{F}(\mathbf{x}_s)\) is the actual forward-tree cost-to-come to \(\mathbf{x}_s\), \(\hat{c}(\mathbf{x}_s,\mathbf{x}_t)\) is an admissible edge-cost estimate, and \(\hat{h}(\mathbf{x}_t)\) is the admissible cost-to-go from \(\mathbf{x}_t\) [2508.21549].

Because direct use of this admissible estimate may be too restrictive, MIT* expands it through a reliability term. The reliability parameter is
\[
\gamma := \frac{g_{\mathcal{F}(\mathbf{x}_s)}}{g_{\mathcal{F}(\mathbf{x}_s) + \hat{c}(\mathbf{x}_s, \mathbf{x}_t) + \hat{h}(\mathbf{x}_t)} = \frac{g_{\mathcal{F}(\mathbf{x}_s)}}{s},
\]
the expansion factor is
\[
e_{\gamma} := \sqrt{1 + (1 - \gamma)^2},
\]
and the estimated initial cost is expanded as
\[
s := s \cdot e_{\gamma}.
\]
If an estimated informed set already exists and the planner still cannot find a valid path inside it, MIT* enlarges it again through
\[
s \leftarrow s \cdot e_\gamma.
\]
The paper notes that the symbol \(s\) is reused for both the admissible prior cost and the expanded estimate [2508.21549].

Using this expanded cost, MIT* defines the estimated informed set as
\[
X_{\tilde{\sigma}} := \{ \mathbf{x} \in X_{\textnormal{free}} \mid \Tilde{f}(\mathbf{x}) < s \},
\]
where
\[
\Tilde{f}(\mathbf{x}) := \Tilde{g}(\mathbf{x}) + \Tilde{h}(\mathbf{x}).
\]
For path-length minimization, the paper specializes this to a prolate hyperspheroid. Let
\[
s_\textnormal{min} := \| x_\textnormal{goal} - x_\textnormal{start}\|_2.
\]
Then
\[
X_{\tilde{\sigma}} = X_{\textnormal{free}} \cap X_\textnormal{HES},
\]
with
\[
X_\textnormal{HES} := \{x \in \mathbb{R}^n \mid \|x - x_\textnormal{start}\|_2 + \|x_\textnormal{goal} - x\|_2 < s \}.
\]
The conjugate-diameter term is
\[
\sqrt{s^2 - s_\textnormal{min}^2},
\]
and the ellipse eccentricity is \(s_\textnormal{min}/s\) [2508.21549].

The hyperellipsoid is also written in matrix form as
\[
X_\textnormal{HES} := \{ x \in \mathbb{R}^n \mid (x -x_\textnormal{center})^T\mathbf{R}\mathbf{P}^{-1}\mathbf{R}^T(x -x_\textnormal{center}) < 1\},
\]
where
\[
x_\textnormal{center} := \frac{x_\textnormal{start}+x_\textnormal{goal}}{2},
\]
and
\[
\mathbf{P} := \operatorname{diag}\left(\frac{s^2}{4},\frac{s^2-s_\textnormal{min}^2}{4},...,\frac{s^2-s_\textnormal{min}^2}{4}\right).
\]
Direct sampling uses
\[
x_\textnormal{hes} = \mathbf{R}\mathbf{L}x_\textnormal{ball} + x_\textnormal{center},
\]
where \(\mathbf{L}\mathbf{L}^T\equiv\mathbf{P}\) and
\[
\mathbf{L} := \operatorname{diag}\left(\frac{s}{2},\frac{\sqrt{s^2-s_\textnormal{min}^2}}{2},...,\frac{\sqrt{s^2-s_\textnormal{min}^2}}{2}\right).
\]
For multiple goals \(X_\textnormal{goal}=\{x_{\textnormal{goal},o}\}_{o=1}^m\), the estimated informed set becomes
\[
X_{\tilde{\sigma}}=\bigcup_{o=1}^mX_{\tilde{\sigma},o},
\]
with
\[
X_{\tilde{\sigma},o}:= \{ x \in X_{\textnormal{free}} \mid \|x - x_\textnormal{start}\|_2 + \|x_{\textnormal{goal},o} - x\|_2 < \Tilde{s}_{\textnormal{est},o} \}.
\]
This phased use of \(X_{\tilde{\sigma}}\) and \(X_{\hat f}\) is the precise sense in which MIT* is “multi-informed” [2508.21549].

## 3. Search architecture, adaptive sampling, and reverse guidance

The planner is organized around two asymmetric structures. The forward tree
\[
\mathcal{T_F} = (V_\mathcal{F}, E_\mathcal{F})
\]
contains valid states and valid edges, while the reverse tree
\[
\mathcal{T_R} = (V_\mathcal{R}, E_\mathcal{R})
\]
is lazy and may contain edges that pass through invalid regions. MIT* initializes
\[
X_{\textnormal{sampled}} \gets \{\mathbf{x}_{\textnormal{start}},\mathbf{x}_{\textnormal{goal}}\},
\]
sets \(E_\mathcal{F} \gets \emptyset\), forms \(\mathcal{T_F}=(V_\mathcal{F},E_\mathcal{F})\), and initializes \(s \gets \infty\), \(\gamma \gets \infty\), and \(e_\gamma \gets \infty\). It then repeatedly adds samples through
\[
X_{\textnormal{sampled}} \stackrel{+}{\leftarrow} \textnormal{adaptiveSampler}(),
\]
builds or updates the reverse tree through
\[
\mathcal{T_R}\gets \textnormal{lazyReverseSearch}(),
\]
and performs forward search while
\[
\textnormal{couldImproveForwardSearch}(\mathcal{T_R})
\]
remains true. If returned forward edges collide, MIT* calls
\[
\textnormal{updateEIS}(s,\gamma,e_\gamma)
\]
and
\[
\textnormal{updateLazyReverseSearch}();
\]
otherwise it inserts valid forward edges into \(\mathcal{T_F}\) and later prunes the sampled set [2508.21549].

The adaptive sampler has three region-selection modes. If there is no valid solution and no lazy estimate, MIT* samples the full free space through
\[
x_\textnormal{pre} \leftarrow \textnormal{sampleSpace}(X_\textnormal{free}).
\]
If there is no valid solution but an estimated cost exists, it samples the estimated informed set through
\[
x_\textnormal{pre} \leftarrow \textnormal{sampleEIS}(x_\textnormal{start},x_\textnormal{goal},s).
\]
Once a valid solution exists, it samples the standard informed set through
\[
x_\textnormal{pre} \leftarrow \textnormal{sampleIS}(x_\textnormal{start},x_\textnormal{goal},c_\textnormal{curr}).
\]
This yields a procedural transition from full-space exploration, to pre-solution EIS-restricted sampling, to post-solution informed-set optimization [2508.21549].

After a preliminary sample \(x_\textnormal{pre}\) is drawn, MIT* applies an obstacle-based adaptation. If \(x_\textnormal{pre}\) is valid, it is added directly. If it is invalid, the planner draws a Gaussian-perturbed point \(x_\textnormal{temp}\) at distance scale \(\delta\); if \(x_\textnormal{temp}\) is valid, it is added, thereby densifying near obstacle boundaries. If \(x_\textnormal{temp}\) is also invalid, MIT* searches along the segment \((x_\textnormal{pre},x_\textnormal{temp})\) for the first valid point \(x_\textnormal{crit}\), thereby targeting critical zones such as narrow corridors and wall gaps. The paper labels these cases \(x_\textnormal{pre,val}\), \(x_\textnormal{pre,inv}\), \(x_\textnormal{temp,val}\), \(x_\textnormal{temp,inv}\), and \(x_\textnormal{crit,val}\) [2508.21549].

The corresponding adaptive-sampler density is defined by an obstacle indicator
\[
\rho(x):= \begin{cases}
1 & \textnormal{if } x \in X_\textnormal{obs} \\
0 & \textnormal{if } x \not\in X_\textnormal{obs}
\end{cases}
\qquad\textnormal{with}\qquad Vol(X_\textnormal{obs}) = 1,
\]
a conditional density
\[
\rho(x_\textnormal{temp}|x_\textnormal{pre}):=\rho_\textnormal{gau}(x_\textnormal{temp})\mathcal{B}(x_\textnormal{temp})/\Psi_\textnormal{const},
\]
and
\[
\pi_\textnormal{adapt}(x_\textnormal{crit}) :=  \int_{X}{\rho(x_\textnormal{temp}|x_\textnormal{pre})\rho(x_\textnormal{pre})}dx_\textnormal{pre}.
\]
Since \(x_\textnormal{crit}\) lies on the segment \(\overline{x_\textnormal{pre}x_\textnormal{temp}}\), the paper writes
\[
x_\textnormal{temp} := \xi \cdot x_\textnormal{crit} - x_\textnormal{pre},
\]
with \(\xi \in \mathbb{R}^{+}\), leading to
\[
\pi_\textnormal{adapt}(x_\textnormal{crit}) :=  \int_{X \cap X_\textnormal{obs}} \frac{\rho_\textnormal{gau}(\xi x_\textnormal{crit} - x_\textnormal{pre})\mathcal{B}(\xi x_\textnormal{crit} - x_\textnormal{pre})}{\Psi_\textnormal{const}}dx_\textnormal{pre}.
\]
This suggests that the sampler is explicitly biased toward critical zones generated by invalid preliminary samples [2508.21549].

Reverse guidance is improved by length-related adaptive sparse collision checks. Rather than using a fixed number of sparse checks for every edge, MIT* updates the sparse-check resolution by
\[
\Delta_{\textnormal{sparse},k} := \Omega \cdot \Delta_{\textnormal{sparse,ini}},
\]
and the number of sparse checks on edge \((x_s,x_t)\) by
\[
\Theta_\textnormal{sparse,k} := \left\lfloor \frac{\| x_s - x_t\|_2}{\Delta_{\textnormal{sparse},k} + 1}\right\rfloor.
\]
Longer edges therefore receive more sparse collision checks. When full forward validation fails, the failure both repairs reverse guidance and yields the prior admissible cost used to create or enlarge the estimated informed set [2508.21549].

## 4. Theoretical properties and asymptotic regime

MIT* explicitly claims probabilistic completeness and asymptotic optimality. The probabilistic-completeness statement is
\[
\lim_{k \to \infty} \mathbb{P} (\{V_\mathcal{F}\cup V_\mathcal{R}\} \cap X_{\textnormal{goal}} \neq \emptyset) = 1.
\]
The asymptotic-optimality statement is
\[
\mathbb{P} (\limsup_{q \to \infty} \min_{\sigma\in\Sigma_q} \left\{ c(\sigma) \right\} = c^*) = 1.
\]
The paper describes MIT* as an anytime method, and these guarantees are presented in that incremental setting [2508.21549].

The asymptotic-optimality argument is based on MIT* using the same Choose Parent and Rewire strategies as EIT* together with an RGG radius condition. The required rewiring radius is written as
\[
r(q) > \eta \left(2 \left(1 + \frac{1}{n}\right){\left(\frac{\lambda( X_{\tilde{\sigma} \cup X_{\hat{f}})}{\zeta_n}\right) \left( \frac{\log(q)}{q}\right)}\right)^{\frac{1}{n},
\]
where \(q\) is the number of sampled states, \(\eta>1\) is a tuning parameter, \(\lambda(\cdot)\) is the Lebesgue measure, and \(\zeta_n\) is the unit-ball volume in dimension \(n\). The paper states that this follows by reference to Lemmas 56, 71, and 72 of Karaman and Frazzoli [2508.21549].

In this account, the key theoretical role of the estimated informed set is not to replace the standard asymptotic machinery but to modify finite-time behavior before the first valid path. Once a valid solution exists, MIT* transitions to the standard informed set
\[
X_{\hat{f}} := \{\mathbf{x} \in X \mid \hat{f}(\mathbf{x}) < c_\textnormal{curr}\},
\]
so the planner’s long-run regime remains within the usual informed-RGG framework. A plausible implication is that MIT* is best understood as an asymptotically conventional informed-tree planner with a nonconventional pre-solution focusing mechanism [2508.21549].

## 5. Empirical behavior, benchmark results, and robotic applications

MIT* is evaluated in Planner Arena, PDT, MoveIt, and OMPL against RRT-Connect variants, Informed RRT*, BIT*, ABIT*, AIT*, and EIT*. The primary objective is path length. Reported common settings include an RGG constant \(\eta = 1.001\), rewire factor \(1.2\), batch size \(100\) for all batch-sorted planners, and Gaussian standard deviation \(\delta\) set to \(10\%\) of the maximum extent of \(\mathcal{C}\)-space. Collision-check resolution is \(5 \times 10^{-6}\) in simulation and \(5 \times 10^{-3}\) in real-world experiments [2508.21549].

The simulation benchmarks comprise Flanking Gap (FG), Random Rectangles (RR), Dividing Walls (DW), and Goal Enclosure (GE), with experiments in \(\mathbb{R}^4\), \(\mathbb{R}^8\), and \(\mathbb{R}^{16}\), and for GE also \(\mathbb{R}^2\). Metrics include initial median time \(t^{med}_{init}\), initial median cost \(c^{med}_{init}\), and final median cost \(c^{med}_{final}\). MIT* consistently reduces initial solution time relative to EIT*, with representative results as follows. In FG-\(\mathbb{R}^4\), EIT* reports \(t^{med}_{init}=0.0068\), \(c^{med}_{init}=1.7640\), and \(c^{med}_{final}=0.7350\), whereas MIT* reports \(0.0053\), \(0.5227\), and \(0.7158\), an initial-time improvement of \(22.05\%\). In FG-\(\mathbb{R}^{16}\), EIT* reports \(0.0129\), \(4.4062\), and \(3.1841\), while MIT* reports \(0.0083\), \(4.1485\), and \(3.1171\), an improvement of \(35.65\%\). In RR-\(\mathbb{R}^{16}\), EIT* reports \(0.2094\), \(4.6672\), and \(3.5319\), while MIT* reports \(0.1423\), \(4.3738\), and \(3.3797\), an improvement of \(32.04\%\). In DW-\(\mathbb{R}^{16}\), EIT* reports \(0.0661\), \(6.2456\), and \(3.8538\), while MIT* reports \(0.0389\), \(5.7522\), and \(3.7675\), an improvement of \(41.15\%\). In GE-\(\mathbb{R}^{8}\), EIT* reports \(0.1203\), \(1.5726\), and \(1.3775\), while MIT* reports \(0.0157\), \(1.5686\), and \(1.3676\), an improvement of \(86.95\%\) [2508.21549].

These results are presented as strongest in narrow passages, obstacle-confined spaces, high dimensions, and scenarios where the first feasible path is difficult to obtain. The paper attributes that behavior to the combination of pre-solution estimated informed sets, obstacle-based adaptive sampling, and more reliable lazy reverse guidance through length-related sparse checks. The ablation study supports that interpretation. It compares MIT*-AS (adaptive sampler only), MIT*-EIS, MIT*-SC (sparse checking), and full MIT*. In GE, adaptive sampling helps less because there are fewer narrow-passage bridge opportunities, while EIS helps more by pruning early. In DW, adaptive sampling helps more because narrow passages are abundant. Sparse checking helps more in DW by improving reverse-search precision and reducing costly restarts. Full MIT* performs best, suggesting that the components are complementary [2508.21549].

The real-world evaluations use the DARKO base-manipulator robot in Beer Barrel-ENV, Shelf-ENV, and Kitchen-ENV, each over \(30\) runs. In Beer Barrel-ENV, MIT* achieves \(93.33\%\) success and median cost \(13.2741\), versus EIT* at \(76.67\%\) and \(16.8917\), and AIT* at \(63.33\%\) and \(18.6277\). In Shelf-ENV, MIT* achieves \(80\%\) success and median cost \(9.7145\), versus EIT* at \(60\%\) and \(11.1054\), and AIT* at \(53.33\%\) and \(12.7851\). In Kitchen-ENV, MIT* achieves \(36.67\%\) success and median cost \(15.9380\), versus EIT* at \(26.67\%\) and \(21.2667\), and AIT* at \(20\%\) and \(23.5611\). Two \(0.1\) s dynamic replanning tasks are also reported: in the horizontal-movement block task, MIT* obtains average cost \(5.5329\) and success \(96.67\%\), compared with EIT* at \(6.7306\) and \(90\%\), and AIT* at \(8.8167\) and \(93.33\%\); in the two vertical-movement blocks task, MIT* obtains success \(93.33\%\) and median cost \(7.8718\), compared with EIT* at \(90\%\) and \(9.6934\), and AIT* at \(86.67\%\) and \(10.2138\) [2508.21549].

## 6. Position within informed-tree research and common misconceptions

MIT* belongs to the informed-tree lineage associated with Batch Informed Trees (BIT*) and Adaptively Informed Trees (AIT*), but it is not interchangeable with them. BIT* established the general paradigm of heuristic search over batches of samples interpreted as an implicit random geometric graph, with an informed subset
\[
X_{\hat f} := \left\{x\in X \mid \hat f(x) < c_{\mathrm{best}} \right\}
\]
defined by the incumbent cost and with almost-sure asymptotic optimality in the standard informed-RGG sense [1707.01888]. AIT* extended that framework through an asymmetric bidirectional architecture in which a reverse search estimates a problem-specific heuristic and a forward search performs expensive edge validation, while remaining almost-surely asymptotically optimal [2002.06599]. MIT* inherits the same broad design space—forward/reverse asymmetry, batch-informed planning, lazy reverse guidance, and asymptotic-optimality-style radius conditions—but distinguishes itself by introducing an estimated informed set before the first valid path exists [2508.21549].

The terminology can be misleading because several later planners use similar suffixes without introducing MIT*. “Informed Steiner Tree*” introduces IST* for multi-goal path finding and does not literally use the name Multi-Informed Trees [2205.04548]. “Task and Motion Informed Trees” introduces TMIT* for integrated task and motion planning and likewise does not define MIT* [2210.09367]. Other neighboring variants include Flexible Informed Trees (FIT*), which changes batch-size policy [2310.12828]; Fully Connected Informed Trees (FCIT*), which removes nearest-neighbour structures by searching fully connected sampled graphs [2411.17902]; Just-in-Time Informed Trees (JIT*), which introduces local just-in-time edge and sample refinement [2601.19972]; and Genetic Informed Trees (GIT*), which learns a richer reverse-search heuristic via reinforced genetic programming [2508.20871]. A common misconception is therefore that MIT* is merely another name for AIT*, IST*, or TMIT*; the literature summarized here indicates that MIT* is a distinct algorithmic proposal centered on pre-solution estimated informed sets rather than a renaming of those earlier methods.

Within that broader family, MIT* is most naturally characterized as an anytime, asymptotically optimal, probabilistically complete informed-tree planner whose novel contribution is front-loaded. Its principal intervention occurs before the first feasible solution, when standard informed planners lose their defining restriction and must resample globally. By deriving a prior admissible cost from failed lazy reverse-search attempts, expanding it via a reliability term, and sampling within the resulting estimated informed set, MIT* turns the pre-solution phase into an informed phase as well. This suggests a specific interpretation of its historical role: not a replacement for BIT*- or AIT*-style informed planning, but a refinement of the stage at which informed restriction first becomes available [2508.21549].

Source: https://www.emergentmind.com/topics/multi-informed-trees-mit