PaperlayerView exact revision on arXiv

Source-linked AI summary

Robotic Guide Dog: Leading a Human with Leash-Guided Hybrid Physical Interaction

Anxing Xiao, Wenzhe Tong, Lizhi Yang, Jun Zeng, Zhongyu Li, Koushil Sreenath

arXiv:2103.14300v2cs.ROcs.SYeess.SY

TL;DR

Robotic guides with large bases or rigid arms have limited ability to operate in narrow spaces. This paper uses a quadruped, a leash, and hybrid interaction modeling with mixed-integer planning to guide a person through such spaces. Experiments report successful collision-free guidance with taut-slack transitions.

  • Problem

    Large bases and actuated rigid guiding arms limit prior robotic guide systems in narrow and cluttered environments.

    Source: §I · p. 1

  • Method

    A Mini Cheetah uses a leash and hybrid interaction model with mixed-integer planning to guide a person while accounting for taut and slack modes.

    Source: B. Contributions · p. 2

  • Results

    Three adults reached their given goals without collision in the narrow-map experiments, and an example task demonstrated leash mode switching.

    Source: B. Robot Guiding Human Experiments · p. 6

  • Takeaways & Limitations

    Experiments indicate the system can guide a person through a narrow space with obstacle avoidance maneuvers and hybrid state transitions.

    Source: §VI · p. 6

Abstract

from arXiv · show

An autonomous robot that is able to physically guide humans through narrow and cluttered spaces could be a big boon to the visually-impaired. Most prior robotic guiding systems are based on wheeled platforms with large bases with actuated rigid guiding canes. The large bases and the actuated arms limit these prior approaches from operating in narrow and cluttered environments. We propose a method that introduces a quadrupedal robot with a leash to enable the robot-guiding human system to change its intrinsic dimension (by letting the leash go slack) in order to fit into narrow spaces. We propose a hybrid physical Human-Robot Interaction model that involves leash tension to describe the dynamical relationship in the robot-guiding human system. This hybrid model is utilized in a mixed-integer programming problem to develop a reactive planner that is able to utilize slack-taut switching to guide a blind-folded person to safely travel in a confined space. The proposed leash-guided robot framework is deployed on a Mini Cheetah quadrupedal robot and validated in experiments.

I. INTRODUCTION

The paper proposes a leash-guided quadrupedal robot to help visually impaired people navigate narrow spaces, where slack-taut switching changes the human-robot system’s dimensions. It combines a hybrid interaction model and mixed-integer planning, then validates the framework on a Mini Cheetah.

I. INTRODUCTION

  • Guide-dog training is individual, time- and labor-intensive, and difficult to scale, motivating an autonomous robotic guide dog.Source: §I · p. 1

1) Robotic Guide Dog:

  • Prior robotic canes using actuated rigid arms may get stuck where the arm cannot fit and require extra control and mechanical-design effort; walker systems’ large bases make narrow-space maneuvering and guidance impossible,,, –.Source: 1) Robotic Guide Dog · p. 1

2) Hybrid System Planning:

  • Prior hybrid-planning work does not address physical human-robot interaction with mobile robots, leaving hybrid path planning for this setting open, –.Source: 2) Hybrid System Planning · p. 2

B. Contributions

  • The proposed leash-guided framework combines a validated hybrid robot-leash-human model with mixed-integer planning over taut and slack modes.Source: B. Contributions · p. 2
  • The framework is deployed on a quadrupedal Mini Cheetah and experimentally validated for collision-free guidance in narrow environments using hybrid mode switches.Source: B. Contributions · p. 2

II. PHYSICAL HUMAN ROBOT INTERACTION (PHRI) MODEL

The pHRI model represents the robot-human configuration in both leash modes. Its degrees of freedom change when the leash switches between taut and slack.

II. PHYSICAL HUMAN ROBOT INTERACTION (PHRI) MODEL

A. Human-robot System

  • A taut leash fixes the separation at l0 and yields four degrees of freedom; a slack leash permits l < l0 and yields five.Source: A. Human-robot System · p. 2

B. Hybrid Dynamic Model

The model assumes a taut leash guides the human along the leash force, while a slack leash leaves the human unmoved. It represents the interaction in the dynamics and switches modes when the leash becomes taut or slack.

1) Taut Mode:

  • When taut, the human is assumed to move in the direction of leash force, with generalized coordinates q and robot velocity and angular-velocity inputs.The robot commands can be sent directly to the Mini Cheetah and tracked by its controller.Source: 1) Taut Mode · p. 2
  • The discrete model switches from slack mode Ss to taut mode St when the leash becomes taut, and returns to Ss when it becomes slack.The transitions are represented in discrete time by ∆s→t and ∆t→s.Source: Figure 3 · p. 3
  • Four discount coefficients, each below one, represent unknown human interaction that can disturb robot tracking when the leash snaps taut.The coefficients can be tuned for prediction; the leash-taut transition interaction and inelastic collision are described as hard to model.Source: 1) Taut Mode · pp. 2–3

2) Slack Mode:

The slack-mode dynamics are defined for the condition in which the cable is slack. In this mode, only the robot moves, and the stated dynamics use geometric relations.

2) Slack Mode:

  • The slack-mode dynamics are denoted ˙qs = fs(qs, us).The passage introduces their definition but does not provide the equations themselves.Source: 2) Slack Mode · p. 3
  • In slack mode, only the robot is moving.This is the physical interpretation given for the right-hand mode in Fig. 3.Source: 2) Slack Mode · p. 3
  • The stated relations (4d) and (4e) come from geometry.The excerpt does not specify the relations' individual forms.Source: 2) Slack Mode · p. 3

C. Leash Tension Model

The leash-tension model uses a linear regression between robot speed and leash tension. The hybrid model distinguishes taut and slack modes, with switching informed by leash direction and tension.

C. Leash Tension Model

  • A linear regression maps robot speed to leash tension by minimizing mean squared error.The model is introduced to connect generalized coordinates to leash tension for physical interaction during path planning.Source: C. Leash Tension Model · p. 3
  • Robot speed projected along the taut leash direction is expressible from generalized coordinates.The passage says this relationship is demonstrated by experiments in Section V.Source: C. Leash Tension Model · p. 3

D. Hybrid Modes Transition

  • A taut leash is treated as analogous to a rigid arm, which is infeasible in confined spaces because it increases the human-robot system's size; the model therefore allows slack mode.Source: D. Hybrid Modes Transition · p. 3
  • Taut mode St requires el · eB ≥ 0 and F ≥ ¯F, while slack mode Ss applies when el · eB ≤ 0 or F ≤ ¯F.The text identifies ¯F as the lower bound representing maximum intrinsic leash tension when slack, and says el · eB < 0 means robot and human will approach next step.Source: D. Hybrid Modes Transition · p. 3

III. PATH PLANNING WITH HYBRID PHYSICAL HUMAN ROBOT INTERACTION

The planner formulates a mixed-integer collocation problem over a horizon from current to target coordinates. Its binary mode variable represents taut or slack cable states, and the problem is implemented in CasADi and solved with IPOPT.

III. PATH PLANNING WITH HYBRID PHYSICAL HUMAN ROBOT INTERACTION

  • The planner formulates a mixed-integer collocation problem over horizon N, using current coordinates and a target from a global planner.The time step is ∆t = t/N.Source: §III · p. 3
  • The binary variable s_k selects the hybrid mode: one denotes taut cable and zero denotes slack cable.The formulation also imposes state, input, and force constraints.Source: §III · p. 3
  • The nonlinear collocation problem is formulated in CasADi with mode-switching logic and solved using IPOPT.Source: §III · p. 3

A. Cost Function

The cost balances target tracking, travel time, and control effort while shaping leash tension; leash-force bounds incorporate the data-driven interaction model into planning.

A. Cost Function

  • The objective combines a soft terminal target penalty and travel-time cost, while stage costs penalize control effort and leash tension changes.The target is not a hard constraint, preserving optimization feasibility; the tension-change cost encourages smooth force variation.Source: A. Cost Function · pp. 3–4
  • The leash-length cost favors more taut modes, which helps guide the person because they are immobile when the leash is slack.The cost also supports faster optimization convergence.Source: A. Cost Function · p. 4

B. Data-driven Leash Tension Constraint

  • Rather than enforce the data-driven leash-tension mapping as an equality, the planner imposes lower and upper force bounds.The two-sided constraint provides greater feasibility than an equality constraint and incorporates physical human-robot interaction.Source: B. Data-driven Leash Tension Constraint · p. 4

C. Obstacle Avoidance

The planner constrains both robot and human clearance from obstacles, then combines a search-based global planner with a local planner that accounts for physical interaction.

C. Obstacle Avoidance

  • Obstacle avoidance uses signed-distance constraints for both the robot and the guided human, with a safety margin for state-estimation and tracking errors.The margin is defined with respect to the robot, human, and obstacle dimensions.Source: C. Obstacle Avoidance · p. 4

A. Framework

  • The end-to-end framework pairs a search-based A* global planner with a collocation-based local planner that incorporates physical human-robot interaction.This framework is designed to safely guide a visually blind person through a cluttered environment.Source: A. Framework · p. 4

B. Global Planner

A reduced-coordinate A* planner generates collision-free waypoints for local planning; onboard sensing tracks the robot and person, while Mini Cheetah’s controller converts velocity commands into joint torques.

B. Global Planner

  • A* searches a grid over reduced coordinates (x, y, φ), then produces a continuous collision-free trajectory with five-dimensional waypoints for the local planner.The heuristic uses approximate dynamics assuming a taut leash and a cosine function of heading to address angular singularity.Source: B. Global Planner · p. 4

C. Robot Localization And Human Tracking

  • Robot localization uses a lidar occupancy map and AMCL, while a depth-RGB camera, OpenVINO detection, and a Kalman filter track the guided person.A two-degree-of-freedom gimbal keeps the person visible across relative orientations.Source: C. Robot Localization And Human Tracking · pp. 4–5

D. Velocity Tracking Controller on Mini Cheetah

V. EXPERIMENTS AND EVALUATION

Experiments report low robot-position prediction error for the identified human-robot model and validate a linear leash-tension model, with 76% of measured forces inside its regression-based bounds.

1) Human-robot dynamic model:

  • The taut-mode model parameters were identified from blind-folded participants’ trajectories, selecting α = [0.8, 0.8, 0.6, 0.8] to minimize prediction error.The model was then incorporated into later guiding experiments; algorithms ran on an onboard Intel computer, with velocity tracking on Mini Cheetah’s real-time computer.Source: 1) Human-robot dynamic model · p. 5, §V · p. 5

2) Leash Tension Model:

  • 76% of measured leash-tension data fell within the linear regression model’s bounds, defined using a 15.06 standard deviation.The linear model was chosen because higher-order models showed no superior prediction performance and would add optimization complexity.Source: Figure 6 · p. 5, 2) Leash Tension Model · p. 5

1) Human-robot dynamic model:

  • 0.023 m average robot-position prediction error and 0.176 m human prediction error were reported for the identified taut-mode dynamic model.The authors describe the human error as acceptable given noise in human detection and estimation.Source: 1) Human-robot dynamic model · p. 5

2) Leash Tension Model:

  • For tension-model validation, the robot guided a person during random movement with the leash slack or taut, recording force, system states, and control commands.Source: 2) Leash Tension Model · p. 5

B. Robot Guiding Human Experiments

In narrow-space experiments, the Mini Cheetah guided blindfolded participants to goals without collisions, using slack–taut leash switching to pass a narrow doorway. The conclusion reports a deployed hybrid-pHRI pipeline and identifies more complicated human-behavior modeling as future work.

B. Robot Guiding Human Experiments

  • 1.0 m was the narrowest region, versus 1.6 m for the taut system; the robot slackened the leash and paused the human while reconfiguring, then resumed taut guidance.The example traversal took roughly 75 seconds.Source: B. Robot Guiding Human Experiments · pp. 5–6
  • Three adults participated, and the fully autonomous human–robot system reached the given goals without collision in the narrow map.The experiments used different initial and goal locations.Source: B. Robot Guiding Human Experiments · p. 6

VI. CONCLUSION AND FUTURE WORK

  • The Mini Cheetah deployment combines a data-driven interaction-force model, hybrid dynamics, and mixed-integer local planning; experiments indicate safe, efficient guidance with obstacle avoidance and mode transitions.Future work will model human behavior with force traction in more complicated ways.Source: §VI · p. 6

B. Robot Guiding Human Experiments

Related papers

Loading 2103.14300v2…