Heuristic Motion Space (HMS)

• Heuristic Motion Space (HMS) is a structured motion representation that biases motion domains using semantic, geometric, probabilistic, and learned heuristics.
• It encompasses methods such as hierarchical motion-memory, state lattice pruning, and learned search heuristics across underwater, generative human motion, and multi-robot planning.
• Empirical results demonstrate that HMS reduces search complexity and improves motion fidelity by integrating multiple heuristic operators into the motion generation process.

Searching arXiv for the cited HMS-related papers to ground the article and verify metadata. arXiv search query: "(Iyer et al., 18 Jul 2025) Heuristic Motion Space (Buchholz et al., 18 Jul 2025, Wen et al., 2020, Bernhard et al., 2021, Attali et al., 2022, Xie et al., 2023, Adabala et al., 2023)" Heuristic Motion Space (HMS) denotes a structured motion representation in which heuristic information shapes how motion is generated, searched, sampled, or reused. In current literature, the term appears explicitly in a hierarchical motion-memory layer for underwater manipulation, and it also functions as an interpretive framework for heuristic-guided state lattices, guiding spaces, learned search heuristics, and generative human-motion representations (Buchholz et al., 18 Jul 2025, Wen et al., 2020, Attali et al., 2022, Bernhard et al., 2021, Iyer et al., 18 Jul 2025). Across these formulations, the common pattern is that an underlying motion domain—configuration states, trajectories, latent motion tokens, or sampled paths—is not treated as a neutral space: it is reduced, biased, or metrized by heuristics that privilege particular regions, directions, or motion primitives.

1. Conceptual scope and terminological usage

In the explicit underwater formulation, HMS is a hierarchical layer on top of a dense PRM, implemented as a repository $\mathcal{H}$ of motion primitives $\{M_i\}_{i=1}^n$, each associated with an uncertainty estimate $U_i$, stored paths $\Pi_i$, and probabilistic reuse information. The planner selects HMS “highway nodes” and cached paths to shorten subsequent searches and updates their probabilities online (Buchholz et al., 18 Jul 2025). In the guiding-space formalism, a related abstraction is defined as a pair

$$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$

where $S$ is a guiding space, $f$ is a projection from configuration space, and $h$ returns a biased sampling distribution over $C$ (Attali et al., 2022). In heuristic-guided lattice planning, the effective motion space is the subset of motion primitives that remain after heuristic pruning at each state, rather than the full lattice action set (Wen et al., 2020). In experience-based heuristic search, the heuristic is learned over motion primitives through a Deep Q-Network and then converted into a cost-to-go estimate for Hybrid A* (Bernhard et al., 2021). In generative human motion simulation, the realized motion space consists of LLM-refined task text, motion tokens in a VQ-VAE codebook, continuous joint trajectories, and evaluation metrics such as MPJPE, PA-MPJPE, and DTW (Iyer et al., 18 Jul 2025).

Taken together, these formulations suggest that HMS is less a single standardized object than a family of constructions in which heuristics alter the operative geometry of motion. The heuristic may be semantic, geometric, probabilistic, learned, or information-theoretic; the motion space may be continuous, discrete, hybrid, or latent.

The acronym itself is overloaded in adjacent literatures. In the cited works, “HMS” also abbreviates human motion simulation (Iyer et al., 18 Jul 2025), human motion segmentation (Dimiccoli et al., 2021, Meng et al., 26 Jun 2025), and Human Mental Search (Mousavirad et al., 2021). In motion-planning and motion-representation discussions, however, “Heuristic Motion Space” refers specifically to heuristic structuring of an underlying motion domain rather than those separate expansions.

2. Mathematical structure of HMS

Different HMS realizations begin from different base spaces, but each paper makes the underlying representation explicit.

Setting	Underlying motion space	Heuristic structure
Generative human motion	$\mathbf{M} \in \mathbb{R}^{T \times 22 \times 3}$, VQ-VAE codebook $\{M_i\}_{i=1}^n$0	LLM prompt rewriting, normalization, MPJPE/PA-MPJPE/DTW
State-lattice navigation	$\{M_i\}_{i=1}^n$1, primitives $\{M_i\}_{i=1}^n$2	$\{M_i\}_{i=1}^n$3, direction field, primitive pruning
Guided sampling	$\{M_i\}_{i=1}^n$4 with projection to $\{M_i\}_{i=1}^n$5	$\{M_i\}_{i=1}^n$6 with $\{M_i\}_{i=1}^n$7
EBHS	Motion-primitive MDP with $\{M_i\}_{i=1}^n$8	$\{M_i\}_{i=1}^n$9
Multi-robot RRT	Joint state space up to 65D	Learned steer $U_i$0 and distance $U_i$1
Underwater PRM-HMS	PRM graph plus repository $U_i$2	Highway nodes, cached paths, BN probabilities

In generative human motion simulation, the kinematic motion representation is

$U_i$3

with joint positions $U_i$4. After normalization and temporal resampling, AI-enhanced, human-prompted, and ground-truth trajectories are embedded in $U_i$5. At the latent level, MotionGPT uses a VQ-VAE tokenizer with encoder $U_i$6, decoder $U_i$7, and codebook $U_i$8, with quantization

$U_i$9

This yields a discrete latent motion space of motion tokens and a continuous decoded joint-space trajectory (Iyer et al., 18 Jul 2025).

In heuristic-guided state lattices, the motion space is a graph over discretized $\Pi_i$0 states, with edges given by motion primitives $\Pi_i$1. The effective HMS at state $\Pi_i$2 is the subset of primitives whose direction aligns with a 2D heuristic direction $\Pi_i$3 within threshold $\Pi_i$4, together with three always-retained basis primitives: “taking a step forward,” “turning in place left,” and “turning in place right” (Wen et al., 2020).

In EBHS, the relevant space is an MDP over motion primitives. With sparse reward

$\Pi_i$5

the learned action value yields a heuristic

$\Pi_i$6

The heuristic is therefore not a geometric distance but a transformed estimate of remaining steps in motion-primitive space (Bernhard et al., 2021).

In joint-space multi-robot planning, the base space is the joint kinodynamic state of all robots under double integrator dynamics, reaching 65 dimensions for 16 robots. The HMS interpretation is induced by a learned distance/cost-to-go field $\Pi_i$7 and a learned steering field $\Pi_i$8, both computed from local observations and composed centrally in the joint state space (Xie et al., 2023).

The guiding-space framework abstracts all of these by separating a reduced space $\Pi_i$9 from the original configuration space $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$0. What makes $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$1 an HMS-like object is not only its lower dimensionality or altered constraints, but the fact that $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$2 turns structure in $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$3 into a biased distribution over $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$4 (Attali et al., 2022).

3. Heuristic operators and mechanisms

The decisive feature of HMS is the mechanism by which heuristic information reshapes motion generation or search.

In generative human motion simulation, the first heuristic is semantic. GPT-4 rewrites free-form task descriptions into motion-aware prompts aligned with MotionGPT’s training vocabulary, particularly HumanML3D action phrases such as “walks at a normal pace,” “throws forward,” and “applies paint using repetitive strokes.” This semantic projection biases inference toward regions of text space from which MotionGPT has learned valid text-to-motion mappings. A second heuristic layer appears in preprocessing: root-centering, scale normalization, axis alignment, and temporal filtering constrain trajectories to a canonical motion space before comparison. A third layer is evaluative: MPJPE, PA-MPJPE, and DTW define scalar costs over trajectories and thereby induce a metric structure on motion quality (Iyer et al., 18 Jul 2025).

In heuristic-guided lattice planning, the central mechanism is directional pruning. A 2D Dijkstra search produces a predecessor direction $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$5 for each cell. During 3D expansion, every candidate primitive has a geometric direction $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$6. If the deviation between $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$7 and $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$8 exceeds $$(f: C \longrightarrow S,\; h: S \times S \longrightarrow \Delta(C)),$$9, the primitive is pruned. The heuristic therefore acts structurally on the action set, not merely on node ordering. The result is a heuristic-restricted motion graph rather than ordinary A* on the full lattice (Wen et al., 2020).

In EBHS, the heuristic comes from experience. A DQN evaluates all outgoing actions at a planning node; those $S$0-values are analytically mapped into child-node heuristics through the sparse-reward relation $S$1. Hybrid A* still performs explicit search and collision checking, but node expansion is biased by learned value estimates over motion primitives rather than by a purely geometric heuristic (Bernhard et al., 2021).

In automated parking, two heuristics are run concurrently within Shared Multi-Heuristic A*: a Reeds–Shepp heuristic that models non-holonomic constraints without obstacles, and a 2D Dijkstra heuristic that models obstacles without orientation constraints. The motion space is therefore viewed through two distinct admissible relaxations, and the shared $S$2-values allow each heuristic to propagate improvements discovered by the other (Adabala et al., 2023).

In joint-space multi-robot planning, the search tree is guided by decentralized, data-driven heuristics. The distance heuristic $S$3 re-ranks nearest-neighbor candidates, while the steer heuristic $S$4 replaces approximate steering with learned accelerations. This induces both a heuristic metric and a heuristic flow field over the same joint motion space (Xie et al., 2023).

In underwater manipulation, the heuristic is explicitly probabilistic. For motion primitive $S$5, goal $S$6, and environment state $S$7,

$S$8

Low-uncertainty, goal-proximal primitives therefore receive higher posterior weight. Because these probabilities are updated through a Bayesian Network with real-time sensor data and path outcomes, the HMS becomes context-aware rather than static (Buchholz et al., 18 Jul 2025).

4. Domain-specific realizations

One major HMS realization is generative and human-centered. The G-AI-HMS pipeline operates in seven stages: text prompt input, GPT-4 prompt enhancement, MotionGPT text-to-motion generation, MediaPipe pose extraction from reference video, spatial normalization and filtering, temporal alignment by resampling to $S$9, and evaluation with MPJPE, PA-MPJPE, and DTW. Its task space covers eight industrially relevant motions including walking, throwing, painting, carrying, and sitting. The resulting HMS spans task text, latent motion tokens, and continuous kinematic trajectories, with evaluation closing the loop between generated and observed motion (Iyer et al., 18 Jul 2025).

A second realization is graph-search based. E$f$0MoP uses a three-layer architecture—global path planning, local path optimization, and time-optimal velocity planning—in which the global lattice planner is itself heuristic-restricted by a 2D environment-aware direction field. The global HMS defines the coarse corridor or homotopy class inside which later optimization proceeds (Wen et al., 2020). Automated parking extends this idea through bi-directional SMHA*, hybrid continuous-discrete states, motion primitives derived from a single-track model, and adaptive motion primitive scaling near obstacles (Adabala et al., 2023). EBHS uses a related Hybrid A* substrate, but its heuristic layer is learned from driving experience rather than handcrafted from geometry (Bernhard et al., 2021).

A third realization is high-dimensional learned planning. In joint-space multi-robot planning, the full search occurs in a continuous-time kinodynamic state space with double integrator dynamics, pairwise collision constraints, and obstacle constraints. The learned heuristics are decentralized at the observation level but composed centrally, allowing tree growth in up to a 65-dimensional joint space (Xie et al., 2023).

A fourth realization is memory-based and explicitly named. In underwater manipulation, HMS is a repository of highway nodes and cached motion primitives layered over PRM. For each successful goal $f$1, the framework stores

$f$2

then updates path probabilities across the repository. Planning for later goals is decomposed into partial A* searches to and from an HMS approach node, with fallback to full A* when no suitable cached route exists (Buchholz et al., 18 Jul 2025).

There is also a representation-learning branch, where HMS refers not to path planning but to the construction of a latent motion space suited for segmentation. Graph-constrained data representation learning introduces an auxiliary matrix $f$3, a nonnegative dictionary $f$4, and a coding matrix $f$5, optimizing

$f$6

so that the learned representation preserves local geometry and temporal smoothness while supporting clustering (Dimiccoli et al., 2021). Temporal Rate Reduction Clustering pursues the same aim through coding-rate terms and a temporal Laplacian, learning representations that are temporally consistent and aligned with a Union-of-Subspaces structure (Meng et al., 26 Jun 2025). These works suggest a broader use of HMS as a heuristic latent space for temporal motion organization rather than for explicit control.

5. Empirical behavior and evaluation

Representative reported outcomes show that HMS constructions are typically justified by reductions in search effort, improvements in motion fidelity, or better clustering structure (Iyer et al., 18 Jul 2025, Wen et al., 2020, Bernhard et al., 2021, Xie et al., 2023, Adabala et al., 2023, Buchholz et al., 18 Jul 2025, Dimiccoli et al., 2021, Meng et al., 26 Jun 2025).

Setting	Reported results	Interpretation
G-AI-HMS	AI better in 6/8 tasks by MPJPE, 7/8 by DTW; MPJPE means 0.353 vs 0.375, $f$7, $f$8; DTW means 61.39 vs 67.01, $f$9, $h$0	LLM-enhanced prompts moved generation toward trajectories closer to human motion
E$h$1MoP	Expanded states decreased by an average of $h$2; graph size about $h$3 of baseline; more than three times faster	Heuristic pruning reduced effective lattice size
EBHS	Pure DQN: $h$4 success in UHL; EBHS: 100% success on the same failure samples; median planning time $h$5 lower than Hybrid A*	Learned heuristics improved speed without inheriting RL failure cases
Joint-space multi-robot planning	Plain RRT fails in every instance for 4 robots; learned heuristics plan up to 16 robots in 65D; BOTH improves nodes by $h$6 vs STEER and $h$7 vs DISTANCE in one regime	Heuristic fields mitigate exponential growth in joint-space search
Automated parking	Slot 27: expanded states 2457 vs 73; execution time 47.8 s vs 11.51 s; reverse path length 7.29 m vs 3.29 m	Multi-heuristic guidance reduces search and preserves path quality
Underwater PRM-HMS	At 30,000 samples: PRM + A* $h$8 s, AHMP $h$9 s; PRM-vs-HMS mean absolute joint errors smaller than PRM-vs-RRT across five tests	HMS reuse improves runtime while keeping PRM-like trajectories

The representation-learning branch reports analogous gains, but in clustering rather than control. Graph-constrained representation learning improves ACC and NMI over both unsupervised and transfer-based HMS baselines on Weizmann, Keck, UT, and MAD; for example, on Weizmann it reports ACC $C$0 and NMI $C$1, compared with transfer-baseline ACC $C$2 and NMI $C$3 (Dimiccoli et al., 2021). Temporal Rate Reduction Clustering reports state-of-the-art HMS performance on five benchmark datasets, including Weiz $C$4 ACC/NMI and YouTube $C$5, with further gains when paired with CLIP features (Meng et al., 26 Jun 2025). In these cases, the “motion space” is validated by its ability to make temporal clustering more separable and temporally coherent.

6. Limitations, ambiguities, and future directions

A recurrent limitation is that HMS is not yet a single canonical formalism. Several papers explicitly note that they do not use the phrase “Heuristic Motion Space” and are being interpreted through that lens rather than introducing standardized HMS theory (Wen et al., 2020, Iyer et al., 18 Jul 2025, Attali et al., 2022). This suggests conceptual breadth, but it also means comparisons across papers often involve different base spaces, different heuristic roles, and different optimality criteria.

Specific limitations recur by domain. Generative human motion simulation relies on MediaPipe as ground truth proxy, defines motion kinematically rather than dynamically, and remains challenged by tool interaction and multi-phase movements (Iyer et al., 18 Jul 2025). Heuristic-guided primitive pruning preserves resolution completeness through invariant basis actions, but theoretical optimality is not guaranteed (Wen et al., 2020). EBHS has no formal guarantee of heuristic admissibility or consistency under function approximation, and its experiments do not address dynamic obstacles or multi-agent traffic (Bernhard et al., 2021). Guiding-space evaluation depends on the choice of target distribution $C$6 and requires KL-based density estimation that becomes difficult in high dimensions (Attali et al., 2022). Learned decentralized heuristics for multi-robot planning depend on expert-generated data and are presently specialized to double integrator dynamics (Xie et al., 2023). Automated parking assumes static 2D occupancy maps and low-speed kinematic modeling (Adabala et al., 2023). The underwater HMS validation is limited to precise arm movements and does not yet cover extensive object handling or whole-body manipulation (Buchholz et al., 18 Jul 2025). Temporal Rate Reduction Clustering notes that the Union-of-Subspaces prior is less suitable for micro-scale manipulative actions in temporal action segmentation settings (Meng et al., 26 Jun 2025).

The future directions are correspondingly diverse. In generative human motion, proposed extensions include multimodal fine-tuning with language, motion, and biomechanical feedback, richer prompt vocabularies, hybrid prompting with expert edits, and integration with physics-based models and digital twins (Iyer et al., 18 Jul 2025). In lattice planning, suggested extensions include multiple heuristic paths, adaptive thresholds, and homology-aware recovery modes (Wen et al., 2020). EBHS identifies dynamic strategic planning as a natural next step (Bernhard et al., 2021). Guiding-space analysis points toward black-box learned guiding spaces, hierarchical guiding spaces, and automated design using information-theoretic evaluation (Attali et al., 2022). Multi-robot work highlights broader dynamics classes and generalization across robot-density regimes (Xie et al., 2023). Parking work points toward dynamic environments, partially observable maps, and additional heuristic channels (Adabala et al., 2023). Underwater manipulation explicitly targets whole-body manipulation and more complex object interaction (Buchholz et al., 18 Jul 2025).

A plausible implication is that HMS will remain a unifying concept only if future work makes the heuristic component explicit at the level of representation, search restriction, or probabilistic guidance. The current literature already shows the essential pattern: motion spaces become substantially more useful when heuristics are not external add-ons but intrinsic operators that define which motions are generated, which states are expanded, which trajectories are retained, and which latent structures are considered meaningful.