KEASL: Adaptive Kinodynamic State Lattice

• KEASL is a kinodynamic planning approach that adapts local maneuver sets to balance exploitation and exploration using terrain and dynamic constraint information.
• It encodes velocity, acceleration, roll, and pitch into graph edges, allowing context-sensitive maneuver evaluations that replace fixed primitive sets.
• Empirical studies show that KEASL improves constraint satisfaction and temporal cost optimization compared to traditional state lattice methods.

Searching arXiv for KEASL, EASL, and closely related kinodynamic state-lattice papers to ground the article in published work. KEASL, or Kinodynamic Efficiently Adaptive State Lattice, denotes a family of kinodynamic motion-planning ideas centered on adaptive maneuver representations rather than fixed, globally uniform primitive sets. In one formulation, KEASL is a recombinant motion planning search space based on the Efficiently Adaptive State Lattice (EASL) in which kinodynamic information is encoded in graph edges through velocity, acceleration, roll, pitch, and direction-dependent evaluation for off-road mobile robot navigation (Damm et al., 24 Apr 2025). In a broader formulation, it is an adaptive state-lattice concept in which a planner is given locally generated maneuver sets that explicitly balance exploitation and exploration, potentially via neural inference conditioned on local obstacle and heuristic information (Sivaramakrishnan et al., 2019). Across these formulations, the defining theme is the replacement of static motion-primitive coverage with context-sensitive maneuver generation or evaluation.

1. Conceptual scope

Classical kinodynamic planning is difficult because, as stated in the learning-based formulation, there is often no local planner available and the only primitive to explore the state space is forward propagation of controls (Sivaramakrishnan et al., 2019). The same source argues that random controls preserve desirable asymptotic properties but lead to slow convergence to high-quality trajectories in practice. KEASL addresses this by making the local maneuver set adaptive.

In the learning-based formulation, the curated maneuver set contains one exploitative maneuver and multiple exploratory maneuvers. The exploitative maneuver is selected from candidate trajectories as the one that gets closest to the goal without collision, while exploratory maneuvers are added greedily to maximize dispersion under a trajectory distance metric. The selection criteria are given as

$$\pi_{new} \in \Pi_{cand}\ \text{minimizing}\ h(\pi_{new}(T)),$$

and

$$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$

The resulting curated set is written as $\hat{U} = \{u^0, u^1, \ldots, u^N\}$ (Sivaramakrishnan et al., 2019).

In the terrain-aware formulation, KEASL differs from prior EASL by using a novel encoding of velocity and acceleration constraints and vehicle direction at expanded nodes, and by evaluating roll, pitch, constraints, and velocities associated with poses along each edge during search while keeping the graph recombinant (Damm et al., 24 Apr 2025). This suggests that the term KEASL is best understood not as a single implementation detail, but as an adaptive kinodynamic lattice paradigm whose concrete realization depends on the planning domain.

2. Relation to state lattices and motion primitives

The principal contrast is with the classical state lattice. Classical state lattice approaches use a precomputed, uniformly spaced set of maneuvers tiled across the state space; their advantages are efficient table-driven control selection and guarantees about reachability and coverage, while their drawbacks are lack of adaptivity and limited suitability for high-dimensional or highly dynamic robots (Sivaramakrishnan et al., 2019). KEASL modifies this picture by making maneuver choice depend on local state, environment, or terrain-dependent dynamic limits.

Representation	Maneuver source	Distinguishing property
Classical state lattice	Precomputed offline	Uniformly spaced, table-driven
Learned adaptive maneuver set	Inferred online via neural network	Explicit exploitation-exploration balance
Terrain-aware KEASL	Edge library with online kinodynamic evaluation	Roll, pitch, velocity, acceleration, direction encoding

The learning-based formulation describes KEASL as a generalization of classic state lattices in which, at each node, the local maneuver set is adaptively inferred using learned models (Sivaramakrishnan et al., 2019). The terrain-aware formulation instead begins from EASL and augments the node and edge representations so that kinodynamic feasibility is represented inside the recombinant search space itself (Damm et al., 24 Apr 2025). In both cases, the static primitive table is no longer sufficient: what matters is whether the local maneuver representation encodes the constraints that actually determine executable motion.

A common misconception is that terrain-aware kinodynamic planning can be reduced to a traditional two-dimensional cost map. The terrain-aware KEASL work explicitly states that this is limiting because a two-dimensional cost map is unable to capture the influence of orientation on the roll and pitch estimates from sloped terrain (Damm et al., 24 Apr 2025). The resulting distinction is foundational: KEASL is not merely a denser lattice, but a lattice whose semantics are altered by path-dependent or context-dependent dynamic quantities.

3. State, edge, and constraint representation in terrain-aware KEASL

In the off-road formulation, each node is defined by discrete position $(x,y)$, discretized heading $\psi$, motion direction $d \in \{+, -, 0\}$, history of velocities and velocity constraints from the start to the current node, and attitude and elevation profile (Damm et al., 24 Apr 2025). An edge from node $i$ to node $j$ encodes a path description from the edge library, an attitude elevation profile $\alpha$, a velocity-acceleration constraint profile $\beta$, and a velocity-acceleration profile $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$0.

The attitude elevation profile is computed pose-by-pose along the edge using a quasi-static suspension model. The stated inputs are robot pose, robot kinematics, and a ground elevation map sampled or interpolated at wheel contact points. Terrain height is sampled at wheel locations, and contact points together with the center of gravity are used to solve for roll and pitch. Poses where roll or pitch exceeds threshold are flagged as non-traversable (Damm et al., 24 Apr 2025).

Velocity and acceleration constraints are then defined as functions of vehicle limits, curvature, attitude, and observability. The terrain-aware description specifies four contributors: max speed and max acceleration from vehicle limits; curvature from path shape; roll and pitch; and observability, where unobserved terrain cells impose strict low velocity constraints. The notation

$$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$1

is used for roll- and pitch-dependent velocity limits (Damm et al., 24 Apr 2025). The qualitative constraint structure is asymmetric: high roll yields a sharp velocity limit to reduce tip-over risk, uphill pitch is less restrictive, downhill pitch is strongly constrained, and the tightest constraint is enforced at each pose.

This representation makes edge feasibility depend on the actual terrain-induced vehicle response rather than on geometric traversability alone. A plausible implication is that KEASL shifts a substantial portion of what would otherwise be deferred to a local planner or controller into the global search representation itself.

4. Recombinant search and maneuver generation

A recombinant lattice assumes that different paths arriving at the same node state are equivalent and can be pruned. Kinodynamic constraints complicate this assumption because two paths can reach the same nominal node with different feasible velocities due to different terrain constraints. The terrain-aware KEASL addresses this by indexing nodes by direction rather than by raw velocity, while storing full velocity and constraint profiles from the start node to the current node (Damm et al., 24 Apr 2025). The stated purpose is to preserve recombination where admissible while retaining the path-dependent information needed for feasibility, optimality, and safety.

The edge-expansion procedure is correspondingly richer than in EASL. The described sequence is: determine the path via the edge library; compute the attitude-elevation profile along the edge; compute pose-wise velocity and acceleration constraints $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$2; use bidirectional Eulerian integration to compute a feasible velocity profile $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$3; integrate velocity to obtain edge duration; and prune edges or nodes that violate physical limits (Damm et al., 24 Apr 2025). Edge cost is duration rather than simple geometric length.

The learning-based maneuver-generation formulation approaches the same problem from the opposite direction: instead of augmenting the lattice semantics, it learns a function that generates a small local maneuver set from the current robot state $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$4, a local obstacle map $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$5, and a local heuristic map $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$6 (Sivaramakrishnan et al., 2019). Feature extraction networks $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$7 and $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$8 process the maps, concatenation operators $$\pi_{new} = \arg\max_{\pi \in \Pi_{cand}} \min_{\pi' \in \hat{\Pi}} d_\pi(\pi,\pi').$$9 combine features with state and previously chosen maneuvers, and the controls are generated according to

$\hat{U} = \{u^0, u^1, \ldots, u^N\}$0

$\hat{U} = \{u^0, u^1, \ldots, u^N\}$1

Here $\hat{U} = \{u^0, u^1, \ldots, u^N\}$2 is exploitative and $\hat{U} = \{u^0, u^1, \ldots, u^N\}$3 are exploratory (Sivaramakrishnan et al., 2019).

These two formulations share a common structural claim: efficient kinodynamic planning improves when the local action set is not fixed in advance, but is conditioned on the local geometry of constraints.

5. Computational mechanics and empirical results

The terrain-aware KEASL paper evaluates the method on a Clearpath Robotics Warthog UGV in three non-flat, outdoor environments described as hilly with grass, ditches or fallen trees, and narrow ravines (Damm et al., 24 Apr 2025). For each test, 24 uniformly distributed goal states at fixed distance were used, yielding 2093 planning queries in total. The maps had 0.2 m resolution, maximum velocity was 2 m/s, and maximum acceleration was 2 m/s².

The reported constraint-satisfaction results are sharp. EASL violated velocity constraints in 100% (2093/2093) of cases and violated roll constraints in 17.96%, whereas KEASL satisfied the constraints by design (Damm et al., 24 Apr 2025). When EASL plans were adjusted to satisfy terrain-dependent velocity constraints, KEASL provided a more efficient route in 83.72% of cases. The reported mean plan times were 54.3 ms for EASL and 352.3 ms for KEASL. The reported mean path durations were 14.23 s for EASL, 27.77 s for KEASL, and 33.93 s for adjusted EASL; the corresponding medians were 13.56 s, 25.56 s, and 33.02 s (Damm et al., 24 Apr 2025). The interpretation given in the source is that KEASL globally optimizes temporal cost with full kinodynamic awareness, while EASL’s nominally shorter plans become slower once terrain-dependent velocity constraints are enforced.

A later extension incorporates stochastic controller behavior into the recombinant search space by simulating noisy rollouts under a pure-pursuit controller with zero-mean Gaussian noise added to velocity commands at each 0.1 s timestep (Damm et al., 6 Aug 2025). Three methods are described: Per-Expansion Rollout (PER), Goal-Edge Rollout (GER), and Goal-Edge Graph Revision (GEGR). Validity is defined conservatively: if any rollout collides, the edge or path is unsafe; edge cost is the average rollout duration,

$\hat{U} = \{u^0, u^1, \ldots, u^N\}$4

The abstract states that incorporating stochastic controller sampling into KEASL leads to more conservative trajectories that decrease predicted collision likelihood when compared to KEASL without sampling, and that baseline planning with expanded obstacle footprints yields more comparable predicted collision likelihood but reduces the planning success rate for baseline search (Damm et al., 6 Aug 2025). The detailed description further characterizes GEGR as practical for real-time on-robot use, whereas PER is extremely computationally expensive.

A second common misconception is that uncertainty-aware planning is equivalent to simple obstacle inflation. The stochastic KEASL results argue against that equivalence: expanded obstacle footprints can improve conservativeness, but they also reduce planning success rate for the baseline search (Damm et al., 6 Aug 2025).

6. Position within kinodynamic planning research

The learning-based KEASL direction is explicitly tied to asymptotically optimal kinodynamic planners. The 2019 position statement argues for integrating modern machine learning frameworks with informed and asymptotically optimal kinodynamic planners so that learned local maneuvers can improve per-iteration performance while preserving desirable properties when random sampling is retained (Sivaramakrishnan et al., 2019). The empirical illustration reports that DIRT-FC-All achieved a FirstSolnIters value of 620 in a “Greedy” environment versus 3446 for Random, and a FinalSolnCost of 130.92 in an “Explore” environment versus 149.47 for Random. This supports the narrower claim that adaptive, neural-network-generated maneuver sets can improve first-solution time and solution quality relative to random controls in the tested settings.

Later work situates EASL and KEASL within a broader ecosystem of motion-primitive planners. The db-LaCAM comparison describes EASL or KEASL methods as using adaptive local refinement of the state lattice during planning and dynamic region focusing, whereas db-LaCAM uses an implicit lattice of precomputed motion primitives combined with a discontinuity bound $\hat{U} = \{u^0, u^1, \ldots, u^N\}$5 and MAPF-based coordination (Moldagalieva et al., 7 Dec 2025). The same source emphasizes that traditional state lattices require strict state continuity between successive motion primitives, while db-LaCAM relaxes continuity by allowing bounded discontinuity between successive motions. This comparison does not invalidate KEASL; rather, it locates KEASL on the side of adaptive single-robot or locally refined kinodynamic search, as opposed to discontinuity-bounded multi-robot coordination.

A different neighboring direction is trajectory-manifold planning. “Trajectory Manifold Optimization for Fast and Adaptive Kinodynamic Motion Planning” represents motion generation through a lower-dimensional differentiable manifold using Differentiable Motion Manifold Primitives (DMMP), with online search performed by sampling latent variables rather than traversing a lattice (Lee, 2024). The same comparison notes that KEASL employs a lattice of precomputed state transitions or motion primitives and that its adaptability is bounded by lattice granularity, whereas DMMP generates continuous-time trajectories through a neural manifold. This suggests a useful categorical distinction: KEASL belongs to the adaptive lattice branch of kinodynamic planning, while DMMP belongs to the continuous generative-manifold branch.

Taken together, these works present KEASL as a technically specific response to a recurring problem in kinodynamic motion planning: fixed primitive sets and purely geometric cost maps are often too weak to capture the local structure of feasibility. Whether the mechanism is learned maneuver curation, terrain-dependent edge evaluation, or stochastic rollout revision, the KEASL idea is to make the search space itself reflect the kinodynamic contingencies that determine executable motion.