EXACT-MPPI: Exact Footprint Navigation

• EXACT-MPPI is a navigation framework that uses exact analytic signed-distance evaluations of robot footprints to assess obstacle proximity directly from LiDAR point clouds.
• It embeds these geometric evaluations within an MPPI control loop, leveraging batched computations in JAX on GPU hardware for real-time receding-horizon planning.
• The framework supports various robot morphologies by directly incorporating arbitrary convex or concave footprint geometries, avoiding intermediate mapping or training.

EXACT-MPPI is a local navigation framework that combines exact, analytic signed distances from a robot’s true $2\text{D}$ footprint to obstacle points with Model Predictive Path Integral (MPPI) control to generate safe motion commands directly from LiDAR point clouds, without any occupancy grid, ESDF, or learned distance model (Peng et al., 28 May 2026). In the specific system published under the name “EXACT-MPPI: Exact Signed-Distance Navigation for Arbitrary-Footprint Robots from Point Clouds via Path Integral Control,” the footprint is represented as a simple polygon for general convex or concave planar shapes, with a rectangle-cover specialization for faster evaluation of rectilinear footprints, and all operations are batched in JAX for real-time receding-horizon control on GPU hardware (Peng et al., 28 May 2026). In the broader MPPI literature, however, the adjective “exact” has a second meaning: it can denote the path-integral control law derived from stochastic optimal control, or the deterministic $\beta \to 0$ limit in which MPPI converges to the exact deterministic optimum under smoothness and SOSC assumptions (Homburger et al., 28 Feb 2025). The term therefore denotes both a specific footprint-aware navigation architecture and a broader theoretical lineage within path-integral control.

1. Definition and conceptual scope

The named framework EXACT-MPPI addresses a geometric failure mode of many local planners for ground robots carrying payloads, implements, or attachments whose effective footprint is complex and non-convex (Peng et al., 28 May 2026). The motivating observation is that most local planners simplify the robot footprint with circles, a rectangle, or a convex hull and rasterize sensor data into occupancy grids or distance fields; when clearance is comparable to the footprint geometry, these simplifications eliminate feasible motions (Peng et al., 28 May 2026). EXACT-MPPI instead keeps the effective footprint as a body-frame polygon or union of rectangles and evaluates obstacle proximity directly against local point-cloud observations during MPPI rollouts (Peng et al., 28 May 2026).

The framework is explicitly training-free. It maps local point-cloud observations and sparse guidance directly to motion commands, without any intermediate map representation and without per-platform training (Peng et al., 28 May 2026). A central practical claim is that the same framework deploys on differential-drive, Ackermann, omnidirectional, and hybrid-mode platforms by changing only the footprint description and motion model (Peng et al., 28 May 2026). This suggests that the method’s portability is tied to its geometric and sampling-based structure rather than to a learned representation.

A recurring source of ambiguity is the phrase “exact MPPI.” In theoretical work, MPPI is called exact for the path-integral class of stochastic closed-loop optimal control problems with input-affine dynamics, quadratic control cost, and the noise–cost matching condition, because the MPPI formula recovers the true optimal closed-loop control law when expectations are exact (Homburger et al., 28 Feb 2025). The same literature also characterizes a deterministic “EXACT-MPPI” limit in which, under smoothness and SOSC assumptions, $$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$ (Homburger et al., 28 Feb 2025). A distinct but related line of work interprets standard Gaussian MPPI as an exact EM update for a latent-variable likelihood model, with the standard weighted-mean update emerging as the exact M-step for a Gaussian family (Wang et al., 29 May 2026). EXACT-MPPI, as a system name, does not replace these meanings; it adds a geometric notion of exactness centered on signed-distance evaluation for arbitrary footprints.

2. Footprint geometry and exact signed-distance evaluation

The geometric core of EXACT-MPPI is the representation of the effective footprint $\mathcal{B}_{\mathrm{eff}}$ as a planar set in the robot body frame that includes the chassis and any rigidly attached geometry (Peng et al., 28 May 2026). For general footprints, the method uses a simple polygon

$$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$

with edges $e_b=v_{b+1}-v_b$, without assuming convexity (Peng et al., 28 May 2026). For orthogonal footprints, it uses a rectangle cover

$$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$

where each rectangle is defined by a center $c_j$ and half-extent $s_j$ (Peng et al., 28 May 2026).

Obstacle points are not projected into a grid. Instead, each rollout state transforms the observed obstacle set into the predicted body frame:

$$p_{h,i}^{b,(r)} = R(\theta_h^{(r)})^\top \big(o_i - t_h^{(r)}\big),$$

so the footprint remains fixed in body coordinates while the points move (Peng et al., 28 May 2026). This design removes the need for occupancy grids, costmaps, or learned encoders in the local planner.

For the rectangle-cover specialization, the signed distance to a box uses the standard closed form

$\beta \to 0$0

$\beta \to 0$1

and the rectangle-union distance is

$\beta \to 0$2

The paper states that this evaluator is exact outside the union and sign-correct inside, with penetration magnitude potentially differing in overlapping regions (Peng et al., 28 May 2026).

For arbitrary simple polygons, the method computes the point-to-segment distance for each edge and combines it with a point-in-polygon sign test:

$\beta \to 0$3

$\beta \to 0$4

$\beta \to 0$5

The paper describes this as the exact Euclidean signed distance to the polygon (Peng et al., 28 May 2026).

This geometric design is the sense in which the system title uses “exact.” The signed-distance evaluator reasons about the robot’s true footprint directly, without convex decomposition, inflation, or learned distance approximation (Peng et al., 28 May 2026). A plausible implication is that the method’s accuracy in tight spaces derives less from a new stochastic control law than from preserving fine footprint geometry all the way into the rollout cost.

3. MPPI formulation and control pipeline

EXACT-MPPI embeds the signed-distance evaluator inside a standard receding-horizon MPPI loop (Peng et al., 28 May 2026). At each control cycle, with current pose $\beta \to 0$6, nominal control sequence $\beta \to 0$7, and horizon $\beta \to 0$8, the controller samples $\beta \to 0$9 perturbed sequences

$$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$0

propagates rollouts under a chosen motion model $$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$1,

$$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$2

and accumulates rollout cost

$$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$3

Here $$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$4 is the minimum signed distance between the footprint and all obstacle points at rollout state $$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$5 (Peng et al., 28 May 2026).

The MPPI weighting is the usual Boltzmann form:

$$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$6

followed by the nominal update

$$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$7

Only $$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$8 is executed before the horizon is shifted and the process repeats (Peng et al., 28 May 2026). In this respect, the control update remains within the classical information-theoretic or path-integral MPPI family (Homburger et al., 28 Feb 2025).

Obstacle handling is driven by the minimum signed distance

$$\lim_{\beta \to 0} \tilde U^\star_{\mathrm{MPPI}}(\beta)=U^\star_{\mathrm{det}}$$9

which feeds the obstacle cost

$\mathcal{B}_{\mathrm{eff}}$0

The first term imposes a large penalty if any point intrudes into the footprint, while the second creates a quadratic repulsion zone whenever $\mathcal{B}_{\mathrm{eff}}$1 (Peng et al., 28 May 2026). The framework also tracks an unsafe rollout flag

$\mathcal{B}_{\mathrm{eff}}$2

and augments the cost with a very large penalty $\mathcal{B}_{\mathrm{eff}}$3, so unsafe rollouts are effectively discarded (Peng et al., 28 May 2026). After the update, the nominal trajectory is re-validated against the condition

$\mathcal{B}_{\mathrm{eff}}$4

and the controller executes a safe stop and resets the nominal sequence if validation fails (Peng et al., 28 May 2026).

The motion model is platform-dependent but structurally interchangeable. The paper lists differential-drive, Ackermann, omnidirectional body-velocity, spin-in-place, and parallel sideways motion models, with the rest of the pipeline unchanged (Peng et al., 28 May 2026). This modularity parallels another contemporary MPPI trend: in humanoid whole-body control, the exact path-integral equations are kept unchanged while sampling is biased by a pretrained RL prior and the task is shaped through modular cost terms (Seo et al., 23 Jun 2026). The shared pattern is that MPPI’s core stochastic update remains fixed while application-specific structure enters through rollout models, priors, and costs.

4. Implementation architecture and computational characteristics

The implementation is written in a NumPy-like style and JAX-compiled, with rollout propagation, coordinate transforms, point-to-footprint distance evaluation, reductions over points and primitives, cost accumulation, and MPPI weighting all fused into GPU-executed kernels (Peng et al., 28 May 2026). The paper highlights vmap and pmap for broadcasting over rollouts, horizon steps, and obstacle points, with just-in-time compilation producing reusable fixed-shape kernels (Peng et al., 28 May 2026).

The dominant workload is the batched point-to-footprint evaluation across rollouts and time. Under the reported settings,

• $\mathcal{B}_{\mathrm{eff}}$5,
• $\mathcal{B}_{\mathrm{eff}}$6,
• $\mathcal{B}_{\mathrm{eff}}$7,
• $\mathcal{B}_{\mathrm{eff}}$8,

the controller performs $\mathcal{B}_{\mathrm{eff}}$9 million distance queries per cycle (Peng et al., 28 May 2026). The reported GPU memory footprint for these settings is approximately $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$0 (Peng et al., 28 May 2026).

Several implementation choices are singled out as computationally decisive. The rectangle-cover specialization reduces runtime by approximately $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$1–$$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$2 compared to edge-based polygon evaluation for rectilinear shapes (Peng et al., 28 May 2026). Obstacle sets are padded to a fixed $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$3 and handled through masking, which stabilizes compilation and batching (Peng et al., 28 May 2026). In hybrid-mode operation, each motion mode is treated as a separate MPPI family while reusing the same signed-distance evaluator and obstacle data (Peng et al., 28 May 2026).

This implementation strategy distinguishes EXACT-MPPI from learned distance evaluators. On an RTX 4060 Ti GPU, for $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$4 points, the analytic JAX evaluator is reported as $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$5 faster on a rectangle, $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$6 faster on a trapezoid, $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$7 faster on a sprayer footprint, and $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$8 faster on a double-sided pruner footprint than DUNE, with consistently better scaling as the number of points grows from $$\mathcal{V}=(v_1,\dots,v_B), \qquad v_b \in \mathbb{R}^2,$$9 to $e_b=v_{b+1}-v_b$0 (Peng et al., 28 May 2026). Deployment overhead is also different: EXACT-MPPI requires only specification of a new polygon or rectangle cover and a single JAX recompilation of less than $e_b=v_{b+1}-v_b$1, whereas DUNE requires approximately $e_b=v_{b+1}-v_b$2 hour of training per footprint (Peng et al., 28 May 2026).

A related but conceptually distinct development in the MPPI ecosystem is the reinterpretation of MPPI as EM. In that setting, for Gaussian proposals with fixed covariance, the standard weighted update

$e_b=v_{b+1}-v_b$3

is identified as the exact M-step of a latent-variable maximum-likelihood problem (Wang et al., 29 May 2026). This does not alter EXACT-MPPI’s implementation, but it clarifies that its rollout update is not heuristic at the algorithmic level even though the geometry and safety logic are application-specific.

5. Empirical behavior across platforms and scenarios

The experimental program emphasizes three claims: acceleration of batched distance evaluation over a learned point-to-robot baseline, preservation of feasible motion where convex-footprint planners fail, and robustness under dense static and moving obstacles (Peng et al., 28 May 2026). The framework is demonstrated on a dual-arm differential-drive indoor robot, the AgileX Ranger Mini in Ackermann, parallel, and spin modes, and a Unitree Go2 carrying a rigid bar and modeled with an omnidirectional body-velocity controller (Peng et al., 28 May 2026).

The clearest evidence for the geometric thesis appears in clearance-limited navigation. In a differential-drive T-shaped corridor experiment with Degree of Narrowness $e_b=v_{b+1}-v_b$4, EXACT-MPPI with the explicit T-shape, Convex-MPPI with a convex hull, and NeuPAN with convex hull and DUNE+NRMP all succeed for $e_b=v_{b+1}-v_b$5–$e_b=v_{b+1}-v_b$6, but at $e_b=v_{b+1}-v_b$7 only EXACT-MPPI succeeds, completing in approximately $e_b=v_{b+1}-v_b$8 at approximately $e_b=v_{b+1}-v_b$9 because the convex hull removes the only feasible passage (Peng et al., 28 May 2026). In an omnidirectional L-shape gap test, both EXACT-MPPI and Convex-MPPI succeed up to $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$0, but at $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$1 only EXACT-MPPI succeeds, with a reported speed of $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$2 and time of $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$3 (Peng et al., 28 May 2026). These results are presented as cases where exact footprint modeling extends the feasible regime (Peng et al., 28 May 2026).

In dynamic clutter, the reported pattern is robustness rather than uniformly faster completion. In the IR-SIM dynamic corridor with eight static and two moving obstacles, over $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$4 runs, EXACT-MPPI attains a success rate of $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$5 versus $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$6 for Convex-MPPI, $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$7 for Rectangle-MPPI, and $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$8 for NeuPAN; the mean navigation time over successful runs is $$\mathcal{B}_{\mathrm{eff}}=\bigcup_{j=1}^R \mathcal{R}_j,$$9 for EXACT-MPPI and $c_j$0 for Convex-MPPI (Peng et al., 28 May 2026). In Gazebo with an AgileX Limo carrying extra load, again over $c_j$1 runs, EXACT-MPPI achieves a success rate of $c_j$2 versus $c_j$3 for NeuPAN, with mean successful times of $c_j$4 and $c_j$5, respectively (Peng et al., 28 May 2026). The paper interprets this as evidence that exactly modeling payload geometry within MPPI produces more robust behavior in dense dynamic settings (Peng et al., 28 May 2026).

Hybrid-mode operation on Ranger Mini illustrates a different advantage. In a narrow scenario with $c_j$6, hybrid MPPI with Ackermann, parallel, and spin modes completes in $c_j$7 versus $c_j$8 for Ackermann-only, a $c_j$9 reduction in completion time (Peng et al., 28 May 2026). In a trap-like scenario under parallel motion with $s_j$0 in the sideways direction, EXACT-MPPI escapes in approximately $s_j$1 while Convex-MPPI fails (Peng et al., 28 May 2026).

The reported perception stack remains relatively lightweight. Point clouds are height filtered and downsampled to a fixed obstacle budget such as $s_j$2, and guidance enters as waypoints or poses transformed through SLAM or LiDAR-inertial odometry (Peng et al., 28 May 2026). The paper reports zero observed collisions in successful trials by construction of the safety margin and validation logic (Peng et al., 28 May 2026). This should not be read as a formal proof of safety; rather, it documents the empirical outcome of the validation-based control loop.

6. Exactness in the broader MPPI literature, limitations, and significance

A central misconception is that EXACT-MPPI’s “exactness” refers only to the geometric signed-distance evaluator. In the broader literature, exactness is layered. For closed-loop stochastic optimal control in the path-integral class, MPPI is exact when the system is input-affine, the stage cost has quadratic control penalty, the noise–cost matching condition

$s_j$3

holds, and expectations are exact; in that case, the MPPI control law recovers the true optimal closed-loop solution (Homburger et al., 28 Feb 2025). For deterministic nonlinear discrete-time systems, exactness is asymptotic: with the scaling $s_j$4 and $s_j$5, the MPPI control converges to the deterministic optimum as $s_j$6, with control suboptimality $s_j$7 and value suboptimality $s_j$8 under the stated smoothness assumptions (Homburger et al., 28 Feb 2025). A third notion appears in EM-MPPI, where the update is exact as an EM mapping for the latent-variable likelihood

$s_j$9

and standard Gaussian MPPI is recovered as the exact M-step for a Gaussian family (Wang et al., 29 May 2026).

Within this conceptual landscape, EXACT-MPPI is best understood as a footprint-aware application framework that uses a standard MPPI backbone while redefining the geometric fidelity of the obstacle term (Peng et al., 28 May 2026). It does not claim an exact solution of a stochastic Hamilton–Jacobi–Bellman equation in the sense analyzed in the theoretical literature (Homburger et al., 28 Feb 2025). Nor does it introduce a new MPPI update rule. Its novelty lies in pairing exact footprint geometry with sampling-based predictive control and making that combination practical at real-time rates on commodity GPUs (Peng et al., 28 May 2026).

The limitations identified for the navigation framework are concrete. The method assumes a planar footprint and $$p_{h,i}^{b,(r)} = R(\theta_h^{(r)})^\top \big(o_i - t_h^{(r)}\big),$$0 obstacle points after height filtering; it does not model vertical clearance, overhanging obstacles, or configuration-dependent body geometry (Peng et al., 28 May 2026). Rollouts use kinematic rather than dynamic models and are therefore aimed at low-speed ground robots rather than aggressive dynamics or rough terrain (Peng et al., 28 May 2026). The planner is local and depends on upstream guidance rather than solving global planning (Peng et al., 28 May 2026). Dynamic obstacles are handled as quasi-static within the short horizon, with reactivity coming from frequent replanning rather than explicit motion prediction (Peng et al., 28 May 2026). Computational complexity remains $$p_{h,i}^{b,(r)} = R(\theta_h^{(r)})^\top \big(o_i - t_h^{(r)}\big),$$1 or $$p_{h,i}^{b,(r)} = R(\theta_h^{(r)})^\top \big(o_i - t_h^{(r)}\big),$$2, so denser point clouds or longer horizons increase cost directly (Peng et al., 28 May 2026).

The significance of EXACT-MPPI therefore lies in a specific systems synthesis. It shows that analytic geometry can be embedded inside MPPI at scale, that arbitrary polygonal footprints can be handled without convexification or learned encoders, and that the same control architecture can transfer across multiple robot morphologies without retraining (Peng et al., 28 May 2026). A plausible implication is that the framework is most valuable when the decisive bottleneck is not nominal trajectory optimization itself, but geometric faithfulness under tight clearances and changing payloads. In that regime, preserving the true footprint can matter more than modifying the sampling law.