Refusing Through the Handle: Two-Speed Material Stiffening for Guide Robots That Disobey

A robotic guide dog has to do something a normal robot never does. It must sometimes refuse its owner. If a blind user tugs the handle toward an open stairwell or into traffic, the right response is to resist. Today that resistance is entirely a software decision carried out by the legs. A safety filter stops or diverts the robot, and the rigid handle passes the result to the user's arm as a sudden jerk. The user cannot tell whether the robot is refusing, malfunctioning, or slipping.

We propose to move part of the refusal into the handle itself, and to state the refusal as a change in how the handle feels: how stiff or how damped the link becomes. Two refusal tiers are matched to two material timescales. A slow tier uses shape-memory alloy (SMA). This is a metal that stiffens and contracts when heated, at the cost of seconds of delay. It gradually stiffens the handle as a predicted hazard approaches. A fast tier uses a pre-stressed snap-through element, cocked in advance, that releases stored energy to lock the handle within an emergency deadline. The SMA then re-cocks that element slowly afterward. The decision to refuse still comes from a perception-based safety filter. The material's job is to make that refusal physical, graded, legible by touch, and gentle on the person.

Core Claim

On a physically coupled guide robot, refusal should be delivered at the human-robot interface as a graded change in mechanical impedance, meaning the handle's resistance to motion. It should not be delivered only as a change in the robot's walking command. We claim that a two-tier impedance refusal does three things compared with a software shield driving a rigid handle. It reduces peak jerk on the user's hand during overrides. It lets users recognize refusal faster and more reliably without spoken explanation. It keeps collision avoidance at least as good. A further claim is specific to the material. Most refusal events on a walking guide robot give seconds of warning, so slow thermal actuation can carry the common case. A pre-stored mechanical release is needed only for a small number of emergencies. This last claim can be checked directly from logs, and the hardware design depends on it.

What Exists and What Is Missing

The proposal sits on top of three guide-robot results that each stop short of the handle.

Transforming a Quadruped into a Guide Robot for the Visually Impaired models the human as trailing the robot through a "Delayed Harness" with fitted lag. It places a LiDAR-based action shield, a filter that blocks commands predicted to cause collisions, between the learned policy and the motors. The shield raised collision-free episodes from 0.52 to 0.84. Its handle is passive, and its override is executed only by the legs. We reuse its shield and its harness model. What we change is the interface that carries the override.

Seeing-Eye Quadruped Navigation with Force Responsive Locomotion Control reads leash tugs from joint and IMU history. It reaches 100% intent accuracy for experts and 70% for novices, and it trains a walking policy that falls far less than model predictive control under pushes. It cannot refuse a dangerous tug. This paper motivates our work, and we use its tug estimator as the input channel.

Do Looks Matter? surveyed 100 people. All of them wanted haptic feedback, and they named intelligent disobedience as non-negotiable. The same respondents wanted a 5 to 15 pound device with 8 to 12 hours of battery life. This sets our weight and power constraints.

Two other lines frame the decision logic. The ICRA Roboethics Challenge 2023 proposes refusal through mediation, offering safe alternatives instead of blunt refusal. It identifies a transparency paradox: explaining a refusal aloud can leak private information or cause friction. A non-verbal, touch-based refusal channel is one direct response to that paradox. Evaluating Robustness over High Level Driving Instruction shows that end-to-end agents obey wrong commands. A segmentation-only agent produced 35 unsafe behaviors in 60 trials, against 3 in 60 with auxiliary grounding. That result motivates our "wrong tug" test design. LMDrive warns that heavy language-model arbitrators add latency. For that reason the decision layer here stays a lightweight geometric shield.

On the materials side, the SMA rehabilitation glove supplies the actuation method. It uses pretrained NiTi wire with a pulley that amplifies the wire's small contraction about sixfold, giving 50 mm of stroke in roughly 7 s at 180 g. That is half the weight of a comparable motor glove. It also supplies the key constraint: full cycles take 3 to 8 s. Robotic metamorphosis by origami exoskeletons supplies the "material as logic" idea, in which a stimulus changes what a structure can do without onboard electronics. Its triggers are environmental and one-way, such as heat and dissolving tabs, and it never considers safety or human contact.

None of these papers changes the interface's mechanics in response to a refusal decision. That gap is what we address.

Research Program

We will build a handle module that mounts between the quadruped and a standard guide-dog grip. The module has a compliant joint at its base and two stiffening mechanisms acting on that joint.

The slow tier is a bundle of pretrained NiTi wires routed through a pulley amplifier, as in the glove design. The wires tension a tendon across the joint, so joint stiffness rises continuously with heating current.

The fast tier is a pre-stressed bistable strip held in its compliant state by a latch. Releasing the latch snaps the strip into a locking configuration. We will compare two candidate releases: a thin SMA pin driven by a short current pulse, and a miniature solenoid. Release latency is an open measurement for both, not an assumption.

After an emergency, the SMA bundle re-cocks the strip over seconds. We will also choose and test the unpowered default. A handle that falls back to a moderate, guiding stiffness on power loss is the fail-safe property we most want to verify.

The control layer extends the action shield. Each step, the shield already projects human and robot footprints. We add an estimate of time-to-hazard for the user's commanded direction. This estimate drives a scalar target stiffness: stiffness rises smoothly as time-to-hazard falls below a horizon matched to the measured SMA response. The bistable lock fires when time-to-hazard falls below the emergency deadline. We will also extend the Delayed Harness model with a variable-impedance term, so that simulation predicts both the user's trailing position and the force on the hand.

The first analysis is a hazard lead-time study. We will run the existing shield in simulated indoor scenes built from Matterport3D, the dataset the guide-robot paper trained on. We will log every intervention. For each one we record the time between the first prediction of a hazard and the latest moment a stop must begin. The resulting distribution, compared against our measured SMA stiffening time, says what fraction of refusals the slow tier can handle on its own.

If that fraction is small, the two-tier design is wrong. The project then reduces to a fast-release mechanism, and the material case weakens considerably. We will report this outcome either way.

Baselines

The naive reference is a rigid handle with no shield, replicating the force-responsive setup of Seeing-Eye Quadruped Navigation. The robot follows the estimated tug direction even toward hazards. This baseline sets the floor for blind obedience in our "wrong tug" trials.

The strongest existing method is the LiDAR action shield with a rigid handle and Delayed Harness model from Transforming a Quadruped into a Guide Robot. Its reported results are a collision-free episode ratio of 0.84 and 1.48 collisions per episode, against 0.52 and 3.96 without the shield. It is the right competitor because it already refuses unsafe commands successfully. Our claim therefore has to show added value in how refusal is delivered, not in whether it occurs. We will reimplement it on the same robot, the same scenes, the same sensor noise, and the same shield parameters. Only the handle and the stiffness controller will differ.

Beating this baseline means two things together. Collision-free ratio and collisions per episode must not be measurably worse. Peak hand jerk during overrides must be significantly lower, and refusal-recognition time must be shorter in the user study. We will not claim a safety gain we do not observe.

Three ablations isolate the key ingredients. The first is slow-tier only, with no bistable lock. It tests whether emergencies need the stored-energy release. The second is fast-tier only, a binary lock with no graded stiffening. It tests whether gradation itself helps recognition and jerk. The third is an electromechanical twin. This is a handle in which a small servo-driven variable damper and a solenoid lock implement the identical stiffness policy. It is matched in workspace and tuned with the same budget, and it separates the value of impedance refusal from the value of the specific materials.

Why not just…?

Why not just put an electromagnetic brake or solenoid lock in the handle? This is the most serious threat, and we treat it as a co-equal arm rather than a straw man. Power-off-engaged brakes are standard parts. They act fast, and they would carry our stiffness policy perfectly well. The evidence that smart materials should win is indirect. The glove paper shows a 50% weight and 40% volume saving for SMA over a motor drive in a different device, and the survey imposes strict weight and battery limits. Neither result proves the advantage carries over to a handle. The electromechanical twin settles it. If the twin matches our handle on jerk, recognition, and collisions at comparable mass and energy per refusal, we will report that the contribution is the impedance-refusal policy, not the material. The materials earn their place only if they win on mass, energy, or fail-safe behavior without losing on latency.

Why not just let the shield stop the legs, as it already does? Leg-level refusal already achieves the 0.84 collision-free ratio, and it may be enough for collision counts. What it cannot change is the force path into the person. With a rigid link, every override arrives at the hand as a jerk, and the user must guess what the jerk means. Our rigid-handle shield baseline measures exactly this. If peak jerk and recognition time do not differ between that baseline and our handle, the objection stands.

Why not just buzz a vibration motor in the grip to signal refusal? Vibration is cheap and addresses the transparency paradox. It does not reduce the mechanical jerk, and it does not physically resist a dangerous tug. It also competes with gait vibration, which is already what limits novice tug recognition to 70%. We will add a vibrotactile-cue arm to the user study, using the rigid handle plus the shield. If vibration matches graded stiffening on recognition time and comprehension, the touch-legibility claim reduces to the jerk-reduction claim alone.

Why not just use a permanently soft handle? Constant compliance would soften jerks, but it also blurs the guidance signal the user follows during normal walking. We will test a fixed-compliance handle set to our handle's softest state. Comparison on harness tracking error, using the Delayed Harness RMSE metric (86.3 mm reported for the fitted model), shows whether adaptivity is needed or whether a soft spring suffices.

Proposed Experiments

In simulation, we will run Matterport3D-derived indoor routes with dynamic obstacles. We will inject "wrong tugs": user commands toward walls, drop-offs, or moving obstacles, generated in the spirit of the Wrong Command Benchmark. The metrics are collision-free episode ratio, collisions per episode, unsafe-compliance count per 60 wrong-tug trials, peak and RMS jerk at the simulated hand, and tracking error. Every arm in the Baselines section is run with matched seeds.

On the bench, we will characterize stiffness range, SMA stiffening and relaxation times, bistable release latency for both triggers, energy per refusal, cycle fatigue over repeated snap-and-reset cycles, and module mass. We will compare all of these against the twin.

On hardware, we will mount the handle on a quadruped of the AlienGo or A1 class. Blindfolded sighted participants will be tested first, then blind and low-vision participants under ethics approval, on courses with staged hazards. Measures are refusal-recognition time, correct interpretation of refusal tier, hand force and jerk, collisions, and a Godspeed social-perception questionnaire as proposed in the ICRA challenge. No numbers have been reported for that questionnaire in this setting, so this would be the first measurement.

The decisive checks are three. The lead-time distribution must show that the slow tier covers a meaningful share of refusals; otherwise the two-speed design is dropped. Graded material refusal must beat the rigid-shield baseline on jerk and recognition without losing collision performance. The comparison with the electromechanical twin determines whether the gain belongs to the materials or only to the policy. We will report whichever answer the data give.