Guidability in Systems: Steering & Control

• Guidability is a concept that characterizes a system’s ability to be steered or to steer another agent under informational, dynamical, and representational constraints.
• It spans diverse applications, from assistive navigation and robotics to guiding language model outputs and verifying automata through fair simulation.
• The framework ties intervention channels, compatibility constraints, and outcome criteria together to ensure effective, safe, and contextually relevant guidance.

Guidability is a domain-dependent technical notion used to characterize whether a system can be effectively steered, or can effectively steer another agent, under the informational, dynamical, and representational constraints of its setting. In assistive navigation and robotics, it concerns converting perception into low-interference locomotion guidance for a human or robot body; in language-model control, it concerns whether a concept direction or collaborator-provided reasoning prefix can actually alter generation in a useful way; and in automata theory, it concerns the coincidence of language inclusion and fair simulation (Yan et al., 24 Jun 2025, Rütte et al., 2024, Boker et al., 2024).

1. Domain scope and general structure

Current research uses the term with several distinct formal meanings rather than as a universal scalar. In blind guidance, the central question is whether sensed environmental information can be translated into effective, wearable, low-interference locomotion guidance with low cognitive burden; in latent-space model steering, the question is whether a concept is not merely decodable but causally steerable; in collaborative reasoning, it is whether a model can leverage a stronger reasoner’s partial trajectory; and in formal verification it is whether every language-contained automaton can be matched by fair simulation (Yan et al., 24 Jun 2025, Rütte et al., 2024, Li et al., 7 Oct 2025, Boker et al., 2024).

A related but distinct use appears in visualization design support. “GuidelineExplorer” treats guidability as making guidance accessible, operational, inspectable, and testable during graph-visualization design by structuring actionable guidelines as if–then statements and exposing them through multiple taxonomical perspectives such as visualization decision, graph type, if-type, and task (Guckes et al., 2024). Here, guidance is neither physical nor control-theoretic; it is interactive decision support.

This suggests a recurring abstraction across fields: guidability is usually operationalized through three linked components—an intervention channel, a compatibility constraint, and an outcome criterion. The intervention may be thigh traction, a latent steering vector, a collaborator’s reasoning prefix, a user-supplied waypoint, or a simulation strategy. The compatibility constraint may be gait phase, fluency preservation, stable humanoid locomotion, or history-deterministic online choice. The outcome criterion may be steering RMSE, PNES, Pass@1, collision-free traversal, or the equivalence of inclusion and simulation.

2. Human-centered assistive guidability

In wearable blind guidance, guidability is explicitly framed as a joint perception-and-locomotion problem: the user must not only detect obstacles and know where to go, but also physically stay on a desired walking direction with low cognitive burden. The wearable system in “Implementing blind navigation through multi-modal sensing and gait guidance” combines GPS, 2D LiDAR, camera, IMU, two motors with rotary encoders, and two traction ropes tied to the thighs in a prototype weighing about $2\,\mathrm{kg}$, with a reported prototype cost of about $\%$. Its core mechanism is gait-phase-aware steering: rotary encoders estimate periodic rope-length variation using relations such as $L'_1 = R(\theta) * L_1$ and $L' = R(\theta) * L_1 + L_2$, and motors apply assistance to the appropriate leg during swing phase to induce stride asymmetry. In the reported experiments, device-assisted steering achieved RMSE $1.6036 \pm 1.4898$ versus $4.2258 \pm 8.1429$ for no assist, indoor obstacle bypass probability was $96\%$ versus approximately $90\%$ for a white cane, indoor hallway speed improved by approximately $23\%$, and outdoor travel time was approximately $32.0\%$ less than with the white cane; the study also notes important limitations, including no stair navigation capability, heavy dependence on walking speed, and use of blindfolded young adults rather than blind participants (Yan et al., 24 Jun 2025).

A related interface-centered formulation appears in “See Spot Guide,” which investigates whether the Boston Dynamics Spot Explorer can guide a human partner through nonvisual interaction channels rather than merely navigate autonomously. The system uses a flexible, responsive handle for robot-to-human communication and a voice-based iOS app for human-to-robot commands. All $\%$0 participants completed the tested tasks, including a $\%$1-foot curved indoor route, a $\%$2-foot approach to a staircase, ascent of $\%$3 steps, and descent after a $\%$4-degree handle pivot. The handle was designed to convey forward motion, turning, stopping, stair approach, and descent posture cues, while the app supported commands such as “Spot, go forward,” “Spot, go backward,” and “Spot, go right/left” (Hata et al., 2024).

Embodied language-guided assistance extends this idea from locomotor cueing to route explanation. “Guide-LLM” introduces a text-based topological map for large indoor environments, with nodes at critical turn or action points and route descriptions emphasizing straight paths and right-angle turns so that guidance is easy to follow. In simulation, the full system achieved $\%$5 success in a Large House and $\%$6 in an Office; removing the system prompt reduced navigation success to $\%$7, and removing the path planning module reduced it to $\%$8 and $\%$9, respectively. The same system reported localization error detection success of $L'_1 = R(\theta) * L_1$0, localization recovery of $L'_1 = R(\theta) * L_1$1, and hazard detection counts of $L'_1 = R(\theta) * L_1$2 total detections, $L'_1 = R(\theta) * L_1$3 true positives, and $L'_1 = R(\theta) * L_1$4 false positives (Song et al., 2024).

Dataset construction for assistive guidability is itself treated as a technical problem in “GuideDog,” which defines accessibility-aware guidance through three standards: S1 “Describe the Surroundings,” S2 “Provide Obstacle Information,” and S3 “Provide a Summary and Direction.” The dataset contains $L'_1 = R(\theta) * L_1$5 image-description pairs, including $L'_1 = R(\theta) * L_1$6 silver-labeled samples and $L'_1 = R(\theta) * L_1$7 gold-labeled samples, drawn from $L'_1 = R(\theta) * L_1$8 source videos spanning $L'_1 = R(\theta) * L_1$9 total hours, approximately $L' = R(\theta) * L_1 + L_2$0 million candidate frames, $L' = R(\theta) * L_1 + L_2$1 cities, and $L' = R(\theta) * L_1 + L_2$2 countries. Its GuideDogQA benchmark contains $L' = R(\theta) * L_1 + L_2$3 object-recognition questions and $L' = R(\theta) * L_1 + L_2$4 relative-depth questions. The experiments show that object recognition can be comparatively strong while relative depth remains a major bottleneck; for example, GPT-4o achieved $L' = R(\theta) * L_1 + L_2$5 on object recognition but $L' = R(\theta) * L_1 + L_2$6 on relative depth, while several open models were near or below chance on depth comparison (Kim et al., 17 Mar 2025).

Virtual environments reveal the same emphasis on movement-ready guidance. In social VR, a “Shared Movement” mechanism lets a blind or low-vision user stand within $L' = R(\theta) * L_1 + L_2$7 ft of a guide avatar, hold the trigger, and move wherever the guide moves; the study with $L' = R(\theta) * L_1 + L_2$8 blind and low-vision participants reports mean ratings of $L' = R(\theta) * L_1 + L_2$9 for comfort using the guide, $1.6036 \pm 1.4898$0 for communication effectiveness, and $1.6036 \pm 1.4898$1 for usefulness, while also showing wide variation in preferred initiative, embodiment, and visibility of the guide (Collins et al., 2024). In demanding room-scale immersive VR wayfinding, “Actionable Guidance Outperforms Map and Compass Cues” found that a view-fixed directional arrow outperformed a north-up minimap and a compass under fog, time pressure, and forced route replanning: estimated marginal means for the composite navigation score were $1.6036 \pm 1.4898$2 for arrow, $1.6036 \pm 1.4898$3 for minimap, and $1.6036 \pm 1.4898$4 for compass, with lower workload for arrow than compass and less interface dwell time (Varshney et al., 18 Mar 2026).

3. Physical interaction, comfort, and collaborative manipulation

A substantial robotics literature treats guidability as a property of physical coupling and shared control. “Quadruped Guidance Robot for the Visually Impaired: A Comfort-Based Approach” models the human as an active responder to traction rather than a passive appendage. The elastic rope obeys $1.6036 \pm 1.4898$5, human walking speed along the rope is modeled as $1.6036 \pm 1.4898$6, and a hybrid standing/walking state machine uses force magnitude and force-rate thresholds to trigger transitions. A nonlinear MPC human planner minimizes terminal and path-tracking error together with force smoothing, yaw smoothness, and motor-induced force changes, while a second MPC computes robot motion commands and a force control device tracks desired force magnitude. In real-world tests, the proposed “elastic rope + FCD” system reduced $1.6036 \pm 1.4898$7 from $1.6036 \pm 1.4898$8 with an inelastic rope to $1.6036 \pm 1.4898$9, reduced $4.2258 \pm 8.1429$0 from $4.2258 \pm 8.1429$1 to $4.2258 \pm 8.1429$2, reduced state changes $4.2258 \pm 8.1429$3 from $4.2258 \pm 8.1429$4 to $4.2258 \pm 8.1429$5, and raised the relative comfort index from $4.2258 \pm 8.1429$6 to $4.2258 \pm 8.1429$7 (Chen et al., 2022).

“Transforming a Quadruped into a Guide Robot for the Visually Impaired” formalizes wayfinding as an MDP in which the robot follows high-level human directional cues such as going straight, turning left or right, or stopping while adjusting detailed trajectories to avoid collisions. Its interaction model, the “Delayed Harness,” represents the human as a lagged compliant follower: the robot moves first, the relative offset is perturbed, and then it decays back toward a default offset. This model achieved RMSE $4.2258 \pm 8.1429$8 in its individualized form versus $4.2258 \pm 8.1429$9 for an optimized fixed-harness model and $96\%$0 for a rotating rod. The accompanying action-shielding mechanism predicts the occupied region of the coupled human-robot pair and suppresses unsafe actions; under ideal sensing, collision-free episode ratio improved from $96\%$1 without shielding to $96\%$2 or $96\%$3 depending on training/test settings, with average collisions dropping from $96\%$4 to as low as $96\%$5 (Kim et al., 2023).

The notion extends beyond wayfinding in “Navigation beyond Wayfinding,” which argues that assistive guidance must often place the user at a precise interaction target and then remain coordinated while the user manipulates the environment. The system alternates between a lead mode and an adaptation mode. In lead mode it computes a stop goal

$96\%$6

balancing collision avoidance, distance to the interaction target, and avoidance of the main object’s swept region. In adaptation mode it uses

$96\%$7

and

$96\%$8

to blend user movement, handle input, and goal direction. In the task-level evaluation, the Full System outperformed both the White Cane and a Non-Adaptive baseline most strongly when target localization was precise and ambiguous, especially for the elevator task, where mean completion time was $96\%$9 s for the Full System versus $90\%$0 s for White Cane, with a significant mean difference of $90\%$1 s ($90\%$2); locate time was significantly shorter than the Non-Adaptive system in every task, with $90\%$3 throughout (Cai et al., 15 Mar 2026).

Ungrounded haptic guidance provides a smaller-scale but conceptually similar case. “Holdable Haptic Device for 4-DOF Motion Guidance” uses coordinated 2-DOF tangential fingertip cues at the thumb and index finger to induce perceived translation or rotation of the hand. In a three-part user study with $90\%$4 participants, users moved in the intended direction even before being told the cue meanings, cue discrimination after brief explanation reached $90\%$5 overall, and every cue exceeded $90\%$6 correct (Walker et al., 2019).

4. Guidability of robot planners and control policies

In motion planning, guidability often denotes the ability of a planner to accept sparse, on-demand external guidance without losing autonomy. “Online, interactive user guidance for high-dimensional, constrained motion planning” studies a $90\%$7-DOF humanoid using Multi-Heuristic A$90\%$8. The planner detects stagnation regions either through expansion delay or through heuristic stagnation, then requests a single intermediate configuration $90\%$9 from the user and injects it through a dynamic heuristic

$23\%$0

Across locomotion and ladder-climbing tasks, the system required about $23\%$1 guidances per run; for example, heuristic-based detection in bipedal locomotion yielded planning time $23\%$2 s with $23\%$3 guidances, while heuristic-based ladder mounting yielded $23\%$4 s with $23\%$5 guidances (Islam et al., 2017).

“HyperGuider” defines a related property for path planning in cluttered and multi-terrain quadruped navigation. A human in VR can set start and goal poses, generate virtual obstacles, and edit poses of a global path; these interventions alter the voxel map used by the planner and reduce the amount of corrective learning the global-local planning loop must perform. In the reported user study, human guidance reduced mean learning time from $23\%$6 s to $23\%$7 s, a $23\%$8 improvement, at the cost of a $23\%$9 increase in path length. NASA-TLX-style results reported physical demand $32.0\%$0, temporal demand $32.0\%$1, frustration $32.0\%$2, and overall performance $32.0\%$3 (Babataev et al., 2022).

GuideWalk moves the same idea into end-to-end humanoid control. It inserts explicit velocity guidance $32.0\%$4 from a DWA-based navigation module into a composite teacher-distillation framework, with the locomotion teacher translating these commands into dynamically feasible whole-body actions. The local planner optimizes

$32.0\%$5

subject to feasible velocity sets. In the main obstacle-navigation task, GuideWalk achieved navigation success rate $32.0\%$6, navigation traversal time $32.0\%$7 s, and proximity to obstacles $32.0\%$8 m. Removing navigation guidance dropped NSR to $32.0\%$9; removing the locomotion teacher reduced obstacle NSR to $\%$00 and beam terrain success rate to $\%$01 (Han et al., 9 Jun 2026).

“OmniGuide” generalizes guidability to pretrained vision-language-action policies by reweighting action generation at inference time with differentiable energy terms over predicted Cartesian trajectories. Its core update adds a guidance gradient to the base flow-matching dynamics: $\%$02 The framework instantiates guidance through collision-avoidance signed-distance fields, semantic target attraction from a VLM, and human-demonstration waypoint attraction. In aggregate real-world experiments, OmniGuide increased success rates from $\%$03 to $\%$04 and collision avoidance rates from $\%$05 to $\%$06; in simulation ablations, initialization-only guidance improved success by $\%$07, denoising guidance by $\%$08, and their combination by $\%$09, while collisions were reduced by $\%$10, $\%$11, and $\%$12, respectively (Song et al., 9 Mar 2026).

5. Guidability in LLMs and collaborative reasoning

In representation-engineering work on LLMs, guidability is explicitly distinguished from detectability. “A LLM’s Guide Through Latent Space” defines the problem as whether a concept encoded in hidden states can be reliably steered at inference time by editing those hidden states along a learned latent direction. The paper uses normalized residual stream representations before attention, the first $\%$13 tokens of the last assistant response with default $\%$14, and three linear probe families—logistic regression, difference-in-means, and PCA. Guidance is norm-preserving and controlled by $\%$15. To evaluate the tradeoff between concept elicitation and fluency degradation, the paper defines the absolute concept effect

$\%$16

and the PPL-normalized effect that underlies PNES. The results show strong concept dependence: average best PNES is $\%$17 for Truthful, $\%$18 for Quality, $\%$19 for Humor, $\%$20 for Creativity, $\%$21 for Appropriateness, and $\%$22 for Compliance. A central empirical conclusion is that best detector $\%$23 best guide; for example, logistic regression is often the strongest detector, but DiM or PCA can be the strongest steering method, and appropriateness guidance often collapses into compliance/refusal behavior rather than semantically clean appropriateness (Rütte et al., 2024).

“Off-Trajectory Reasoning” defines Guidability as the ability of a reasoning model to successfully leverage a guiding steer from a stronger collaborator to surpass its solo-reasoning ability. The benchmark selects hard math questions where the tested model solves at most $\%$24 samples on its own, removes the model’s own prefix ($\%$25), and provides the first $\%$26, $\%$27, $\%$28, or $\%$29 of a stronger model’s correct reasoning trace. Across $\%$30 evaluated models, shared-subset Guidability remained under $\%$31; even the best model, R1-Qwen-32B, averaged $\%$32 on the shared subset and $\%$33 on its individual subset. The paper’s most pointed analysis shows that guiding traces already contained the correct answer in $\%$34 of cases on average, yet average individual Guidability was only $\%$35, implying that models often failed even when the solution was effectively present. Reinforcement learning after SFT improved guidability by $\%$36–$\%$37, suggesting that this behavior is learnable but not well induced by standard solo-reasoning training (Li et al., 7 Oct 2025).

These results suggest that guidability in LLMs is stricter than either benchmark strength or linear probe accuracy. A concept may be separable yet not cleanly steerable, and a strong solo reasoner may still be poor at building on a collaborator’s partial trace. In both settings, the crucial issue is whether the intervention aligns with a causal axis of generation rather than merely with a correlated feature bundle.

6. Formal-language meaning: inclusion, fair simulation, and history-determinism

In automata theory, guidability has a precise semantic meaning. An LTS $\%$38 is guidable with respect to a class $\%$39 if it simulates every $\%$40 whose language is included in $\%$41: $\%$42 Here fair simulation is stronger than inclusion, so guidability is exactly the property that inclusion and simulation coincide “from below.” The comparison notion is history-determinism, defined through a letter game in which Eve resolves nondeterminism online from the prefix read so far. History-determinism always implies guidability; the paper’s main question is when the converse holds (Boker et al., 2024).

The answer is a set of sufficient criteria. The paper proves that history-determinism and guidability coincide if the class is closed under determinisation or under language-equivalent history-deterministic representatives, or if the class is closed under suitable 1-token ghost constructions, or if Adam’s winning strategies in the letter game can be recognized by a deterministic automaton whose projection has a 1-token ghost. Using these criteria, the paper shows coincidence for $\%$43-regular automata, fixed-index parity and weak automata, uniform infinite-state classes with safety or reachability conditions including vector addition systems with states, one-counter nets, pushdown automata, Parikh automata, and timed automata with safety or reachability acceptance conditions, as well as visibly pushdown automata with any $\%$44-regular acceptance condition and linear automata. It also shows that the notions differ for Büchi automata with a bounded number of states and for timed automata with a fixed number of clocks (Boker et al., 2024).

The algorithmic significance is substantial. Because direct guidability quantifies over all language-contained automata, it is not easy to decide in general. When guidability coincides with history-determinism, however, deciding guidability reduces to deciding HD. The paper explicitly extracts decidability and complexity consequences for several classes, including PTIME for Büchi and coBüchi automata and EXPTIME for parity automata, safety and reachability timed automata, and visibly pushdown automata (Boker et al., 2024).

7. Recurrent trade-offs and unresolved issues

Across domains, guidability is usually strongest when the steering representation is close to the action interface and weakest when substantial cognitive or algorithmic translation remains. The assistive gait-guidance system explicitly argues against reliance on high-level audio alone and instead intervenes during the correct gait phase (Yan et al., 24 Jun 2025). The immersive VR study similarly shows that a directional arrow outperforms richer but more interpretive cues such as a minimap or compass when users must move under fog, time pressure, and route replanning (Varshney et al., 18 Mar 2026). In graph visualization, GuidelineExplorer makes the same point in a different register: actionable if–then guidance, directly applicable to a graph, is treated as more usable than a diffuse literature of partially contextualized recommendations (Guckes et al., 2024).

A second recurrent theme is that nominal capability is not sufficient. Stronger solo reasoners were not clearly more guidable in off-trajectory reasoning (Li et al., 7 Oct 2025). Better probes were not necessarily better guides in latent-space steering, and appropriateness directions could be highly detectable yet poorly guidable as appropriateness (Rütte et al., 2024). In robot control, locomotion competence without an explicit guidance interface produced oscillations, detours, or unsafe behavior, while guidance without terrain-adaptive execution collapsed on difficult terrain (Han et al., 9 Jun 2026).

A third theme is ecological validity. Several assistive systems were evaluated on sighted but blindfolded participants rather than blind users, which the papers themselves note as a limitation (Yan et al., 24 Jun 2025, Hata et al., 2024). Guide-LLM remains simulation-based and reports $\%$45 true-positive versus $\%$46 false-positive hazard detections, indicating that safety-aware guidance can still be noisy (Song et al., 2024). GuideDog makes the corresponding perception bottleneck explicit by showing that relative depth remains weak even when object recognition is comparatively strong (Kim et al., 17 Mar 2025). In formal verification, the analog of this issue is resource-bounded mismatch: once the class is restricted, as with timed automata with a fixed number of clocks, guidability can separate from history-determinism (Boker et al., 2024).

Taken together, the literature supports a technical, non-metaphorical understanding of guidability. It is not merely whether a system can be influenced, and not merely whether it can provide guidance cues. It is whether steering information can be injected at the right abstraction level, in a form compatible with the target system’s internal dynamics, so that the resulting behavior remains effective under the relevant criteria of the domain—safety, fluency, controllability, simulation, or collaborative reasoning success.