PaperlayerView exact revision on arXiv

Source-linked AI summary

TAPNAV: Humanoid Navigation through Tactile Active Perception

Huaze Liu, Zhenyu Wu, Jaehwi Jang, Junjie Sheng, Andrew Collins, Aaron Xie, Zhaoyuan Gu, Kaijie Zhu, Ding Jiang, Kevin Cai, Ye Zhao

arXiv:2610.10748v1cs.RO

TL;DR

Vision-denied navigation is difficult because conventional visual sensing can be unreliable and humanoid localization must combine sparse tactile observations with motion estimates. TAPNAV plans tactile probes alongside routes and whole-body control; simulation and hardware results report improved navigation success and lower pose error over baselines.

  • Problem

    RGB-D cameras and LiDAR can become unreliable in visually degraded environments, while tactile navigation must select informative contacts and integrate sparse observations with motion estimates.

    Source: §I · p. 1

  • Method

    TAPNAV combines tactile-based pose belief, uncertainty-aware route planning, and local probes selected by expected information gain for vision-denied humanoid navigation.

    Source: §I · p. 2, §III · p. 2

  • Results

    TAPNAV improves navigation success and lowers pose error over ablation baselines; Odometry-only average success was 22.5%.

    Source: A. Simulation Results · p. 7, §VI · p. 8

  • Takeaways & Limitations

    Simulation and hardware experiments demonstrate goal-directed navigation using tactile feedback without vision.

    Source: §I · p. 2, §VI · p. 8

    TAPNAV assumes a known static floor plan and approximates the particle belief as a unimodal Gaussian during planning.

    Source: §VI · p. 8

Abstract

from arXiv · show

Navigation in vision-denied environments is challenging for humanoid robots because proprioceptive odometry drifts and localization uncertainty accumulates rapidly. We present TAPNAV, a tactile active-perception framework that enables humanoid navigation toward a goal by actively probing surrounding structures without relying on vision. TAPNAV maintains a pose belief from odometry, IMU, and tactile contact observations, and couples uncertainty-aware global route planning with information-gain-driven local probing. The global planner searches for routes that keep predicted localization uncertainty bounded by exploiting opportunities for tactile correction, while the local planner selects probe actions that maximize expected information gain. A whole-body controller coordinates the humanoid's locomotion and end-effector contact to execute the planned navigation and probe motions. We evaluate TAPNAV in simulation and on a Unitree G1 across different floor plans and obstacle geometries. TAPNAV achieves lower state estimation error and a higher task completion rate than baselines. These results demonstrate that actively planning physical interactions with the environment can provide localization cues for reliable humanoid navigation without vision.

I. INTRODUCTION

TAPNAV addresses vision-denied humanoid navigation by combining tactile probing with localization and planning. Its route planner accounts for pose uncertainty, while its local planner selects informative contacts for a coordinated navigation system.

A. Tactile Localization and Active Sensing

  • Extending tactile sensing from object-centric tasks to navigation requires informative contacts, whole-body balance, and integration of sparse tactile observations with motion estimates.Prior work also considered information-based interactions and localization from gait-induced contacts without active action selection.Source: §I · p. 1, A. Tactile Localization and Active… · p. 2

I. INTRODUCTION

  • TAPNAV combines uncertainty-aware route planning and information-gain-driven tactile probing with whole-body control for vision-denied navigation.The framework was validated in simulation and real-world office-scale experiments.Source: §I · p. 2

B. Belief-Space Planning and Active Localization

  • Unlike belief-space and active-localization approaches that primarily rely on remote perceptual observations, TAPNAV plans motions that use contact to maintain the robot-state belief.Source: B. Belief-Space Planning and Active… · p. 2

C. Contact-Rich Humanoid Loco-Manipulation

  • Where conventional humanoid loco-manipulation uses contact for manipulation, TAPNAV selects when and where to probe to reduce localization uncertainty.Source: C. Contact-Rich Humanoid… · p. 2

III. SYSTEM OVERVIEW

  • The estimator combines odometry, IMU heading, and tactile observations in a pose belief; the global planner then computes a collision-free route with bounded predicted uncertainty.Source: §III · p. 2, Figure 2 · p. 3
  • At each subgoal, the local planner selects a feasible end-effector probe by expected information gain, and the low-level controller executes the route displacement and probe.Probe directions are discretized from −40° to 110°, subject to arm kinematic feasibility.Source: §III · pp. 2–3, Figure 2 · p. 3

IV. METHOD

TAPNAV’s tactile end-effectors detect environmental contact, and the processing pipeline converts sensor activations into a binary contact observation for the planner.

A. Tactile Sensing and Contact Processing

  • Each end-effector uses a SensX 160 tactile sensor beneath a silicone pad in a curved rigid housing for contact sensing across approach angles.The sensor has a 20 × 8 taxel array, a 20 × 50 mm active area, and a force range up to 100 N.Source: A. Tactile Sensing and Contact… · p. 3
  • After offset correction and filtering, total activation T(t) is converted to estimated normal force and thresholded into a binary contact outcome for the planner.The force conversion uses the inverse exponential response model introduced in WT-UMI.Source: A. Tactile Sensing and Contact… · p. 3, Figure 3 · p. 3

1) Pose belief:

The pose belief is maintained as weighted pose particles propagated by odometry and corrected using tactile and IMU observations. The estimator resamples when particle weights become too concentrated and approximates the belief as a Gaussian for efficiency.

1) Pose belief:

  • The robot’s pose belief is represented by weighted particles, with a Gaussian approximation computed from them for computational efficiency.The Gaussian mean and covariance are calculated from the weighted particles.Source: 1) Pose belief · p. 3
  • Odometry propagates each particle by composing its prior pose with a noisy body-frame increment.The noise model is Q(ut), and ⊕ denotes SE(2) pose composition.Source: 1) Pose belief · p. 3
  • Tactile and heading likelihoods reweight the propagated particles, which are resampled when effective sample size falls below 0.5Np.Small Gaussian roughening is applied after resampling.Source: 1) Pose belief · p. 4

2) Tactile likelihood:

The tactile likelihood compares a probe’s observed stopping point with the mapped surface under each pose hypothesis. It favors hypotheses consistent with contact and penalizes contact-free observations only when they predict the probe should have reached a surface.

2) Tactile likelihood:

  • A tactile probe provides geometric information about pose by relating its observed end-effector stopping position to the known map.The stopping position is transformed into the map frame and compared with the nearest map surface.Source: 2) Tactile likelihood · p. 4
  • The signed normal distance δ(i) is zero on the surface, negative for predicted penetration, and positive when the probe stops short.These cases distinguish surface contact, an inconsistent contact-free prediction, and a probe that does not reach the surface.Source: 2) Tactile likelihood · p. 4
  • Contact observations penalize hypotheses according to their distance from the surface, while contact-free observations are inconsistent only when they predict penetration.The resulting likelihood updates and normalizes particle weights.Source: 2) Tactile likelihood · p. 4

3) Uncertainty-bounded global planner:

The global planner searches for short routes that keep predicted localization uncertainty within bounds by accounting for tactile correction opportunities. At route subgoals, the local planner selects probes for expected information gain, using local probe-to-surface geometry to assess contact and pose uncertainty.

3) Uncertainty-bounded global planner:

  • The constrained A* planner returns the shortest candidate route satisfying position and heading uncertainty bounds, pruning violations and replanning when pose deviation exceeds a threshold.It predicts belief evolution at displacements limited to |∆xt| ≤ 0.4 m and uses surface proximity to estimate tactile covariance reduction.Source: 3) Uncertainty-bounded global planner · p. 4

C. Expected-Information-Gain-Driven Action Selection

  • At each route subgoal, the local planner chooses a tactile probe by maximizing expected information gain across contact and contact-free outcomes.The selection uses the probabilities and information gains of possible outcomes.Source: C. Expected-Information-Gain-Driven… · p. 5

1) Contact geometry of an action candidate:

  • For each candidate probe, the planner estimates the end-effector’s closest mapped surface and approximates it locally as a plane with an outward normal.It linearizes signed normal distance around the current mean pose to relate pose errors to distance changes.Source: 1) Contact geometry of an action… · p. 5

2) Contact probability:

The contact-probability model treats the probe’s signed distance from a local surface at its travel limit as uncertain. It combines pose and control uncertainty to estimate the chance that a candidate probe reaches the surface.

2) Contact probability:

  • The probe margin is the signed normal distance from the end-effector’s travel-limit position to the local surface, with a negative value indicating the surface lies within reach.The margin is modeled as a Gaussian random variable.Source: 2) Contact probability · p. 5
  • The margin uncertainty reflects both pose uncertainty and control noise.The passage identifies the margin variance as induced by these uncertainty sources.Source: 2) Contact probability · p. 5
  • The probability of contact for a candidate action is computed from the Gaussian margin model using the standard normal cumulative distribution function.The resulting probability is denoted pt(a).Source: 2) Contact probability · p. 5

3) Expected information gain:

TAPNAV scores probe actions by expected information gain across both contact and contact-free outcomes. It updates the pose belief after each observation and repeats probing until the best expected gain falls below a threshold.

3) Expected information gain:

  • A contact provides a signed normal-distance measurement, while a contact-free observation truncates the margin distribution at zero.The contact-free outcome indicates that the probe did not encounter the surface.Source: 3) Expected information gain · p. 5
  • The contact-free information gain uses the variance ratio of the truncated normal distribution relative to its prior.The fraction of margin variance attributable to pose uncertainty is also used in the calculation.Source: 3) Expected information gain · p. 5
  • Figure 5(b) shows pose-belief reduction for both contact and contact-free outcomes.Figure 5(a) depicts margin uncertainty before a probe and after either outcome.Source: 3) Expected information gain · p. 5, Figure 5 · p. 5
  • Candidate probes are scored by expected information gain across contact and contact-free outcomes.The resulting tactile observation is fed back into the particle filter, and probing continues until the maximum expected gain falls below a threshold.Source: 3) Expected information gain · p. 5

D. Low-level Controller

The low-level controller coordinates end-effector interaction and footstep placement through upper- and lower-body policies. The policies use shared proprioception and are jointly optimized for stop-and-go navigation with contact.

D. Low-level Controller

  • The controller tracks end-effector pose and force alongside footstep placement to support frequent stop-and-go maneuvers while maintaining contact.The upper-body goal specifies end-effector poses and target normal forces; the lower-body goal specifies footstep placements and timing.Source: D. Low-level Controller · pp. 5–6
  • The upper-body policy takes an upper-body goal, while the lower-body policy takes a locomotion goal, and both produce their respective body actions.The policies use shared proprioception that includes a history of joint positions and prior actions.Source: D. Low-level Controller · p. 6
  • The upper- and lower-body policy parameters are jointly updated with PPO using separate goal-conditioned objectives.Their actions are combined into a whole-body action before it is sent to the low-level PD controller.Source: D. Low-level Controller · p. 6

1) Upper-body Policy Training:

Upper-body policy training targets accurate end-effector pose and force tracking through a two-stage curriculum. It progresses from pose tracking with inverse-kinematics references to force-command tracking using simulated interaction forces.

1) Upper-body Policy Training:

  • The upper-body policy is trained in two stages: end-effector pose tracking first, followed by force-command tracking.The stated training goal is accurate tracking of target end-effector pose and force.Source: 1) Upper-body Policy Training · p. 6
  • In stage one, sampled end-effector poses and posture-regularized inverse kinematics produce reference joint positions for policy tracking.The policy tracks both the reference joint positions and their corresponding end-effector poses.Source: 1) Upper-body Policy Training · p. 6
  • In stage two, a simulated spring-damper field applies force along the probe direction to train force-command tracking.The setpoint is placed so that holding the target pose requires exerting the commanded normal force.Source: 1) Upper-body Policy Training · p. 6

2) Lower-body Policy Training:

The lower-body policy is trained to track rest-to-rest stepping references generated from sampled support configurations and footstep targets. The experiments evaluate TAPNAV in simulation and on hardware, with questions covering state estimation, position adjustment, and real-world navigation.

2) Lower-body Policy Training:

  • Rest-to-rest step references are generated from double-support configurations and footstep targets, with a piecewise-constant ZMP mapped to a CoM trajectory by the 3D-LIPM.The swing foot follows a cubic spline with zero lift-off and landing velocities and fixed swing height.Source: 2) Lower-body Policy Training · p. 6
  • The lower-body policy tracks reference CoM, swing-foot, and joint trajectories.The references are constructed by sampling double-support configurations, footstep targets, and step durations.Source: 2) Lower-body Policy Training · pp. 6–7

V. EXPERIMENTS

  • Experiments use MuJoCo simulation and a Unitree G1, addressing bounded state estimation across maps, position-adjustment accuracy, and real-world navigation reliability.The paper states that the experiments use the parameters in Table I.Source: §V · p. 7

A. Simulation Results

In simulation, TAPNAV uses tactile contacts to reduce localization uncertainty and outperforms active-touch baselines in navigation success; hardware trials demonstrate navigation in an office hallway and kitchen. The conclusion notes that the framework assumes a known static floor plan and unimodal Gaussian planning beliefs.

A. Simulation Results

  • Contacts with different surface normals around corners and pillars reduce position uncertainty, while flat-wall contacts constrain position along the surface normal and can reduce heading uncertainty.Across four MuJoCo layouts, estimated trajectories closely follow simulator ground truth.Source: A. Simulation Results · p. 7
  • TAPNAV achieves a 95% average success rate, versus 77.5% for Random-touch, 38.8% for Sweep-touch, and 22.5% for Odometry-only, with lower position error and comparable yaw error to active-touch baselines.Across added yaw noise, TAPNAV also has the lowest mean position and yaw errors and the highest success rate.Source: A. Simulation Results · p. 7, Table II · p. 8
  • In open-loop goal tracking, the footstep controller follows target sequences with smaller drift than velocity tracking, whose heading errors accumulate across successive targets.The comparison uses 30 trials per policy and the same nominal path.Source: A. Simulation Results · p. 7, Figure 9 · p. 8

B. Real-World Experiments

  • On a Unitree G1, successful trials navigate an office hallway and kitchen, with the robot probing different surfaces while turning and moving around a counter.These experiments evaluate the integrated estimation, planning, and control pipeline under real-world sensing and control uncertainties.Source: B. Real-World Experiments · p. 8

VI. CONCLUSION

  • TAPNAV integrates tactile active probing, belief-space localization, uncertainty-aware global planning, and whole-body control for vision-denied goal-directed navigation.Simulation and hardware experiments demonstrate navigation using tactile feedback.Source: §VI · p. 8
  • The framework assumes a known static floor plan and approximates the particle belief as a unimodal Gaussian during planning.The conclusion identifies multimodal beliefs and partially known maps as planned extensions.Source: §VI · p. 8
Loading 2610.10748v1…