TAPNAV: Humanoid Navigation through Tactile Active Perception

Anonymous Authors

Anonymous Institutions

Under Review

TAPNAV planned path, planned tactile action, and Unitree G1 deployment in a kitchen environment
TAPNAV enables vision-denied humanoid navigation through tactile active perception. The humanoid actively makes contact with the surrounding environment at informative locations, using contact observations to reduce localization uncertainty and achieve reliable navigation.

Abstract

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.

Method

Vision-Denied Navigation

TAPNAV actively probes surrounding structures to localize and navigate toward a goal without relying on cameras or LiDAR.

Humanoid Navigation through Tactile Active Perception

System Overview

TAPNAV localizes the robot on a known, static map using odometry, an IMU-based heading estimate, and tactile contact with mapped surfaces. A global route planner, local tactile action planner, pose estimator, and low-level controller jointly enable goal-directed navigation without vision.

TAPNAV system overview with robot observations, belief update, global and local planners, and low-level controller
Odometry, IMU, and tactile observations together update the robot's SE(2) pose belief. The global planner computes an uncertainty-bounded route with predicted tactile corrections, the local planner selects informative probes based on expected information gain, and the low-level controller executes locomotion and probing.

Tactile Sensing and Contact Processing

Tactile image processing, force calibration, contact outcome, and tactile end-effector construction
Tactile end-effector design and contact-processing pipeline. The tactile image is aggregated into a total activation, calibrated to a normal force, and converted to a binary contact outcome using a force threshold.

Each end-effector integrates a TouchTronix SensX 160 tactile sensor covered by an Ecoflex 00-50 silicone pad. The tactile image is aggregated into a total activation, calibrated to an estimated normal force, and thresholded to produce the binary contact outcome used by the planner.

Low-Level Controller

TAPNAV low-level controller with upper-body and lower-body policies
The low-level controller is composed of upper- and lower-body policies. The upper-body policy tracks target end-effector pose and force, while the lower-body policy tracks footstep commands. Both policies are jointly trained with shared whole-body proprioception.

To support frequent stop-and-go maneuvers while maintaining contact with the surroundings, TAPNAV learns a whole-body controller that tracks end-effector pose and force together with footstep placement.

Experiments

We evaluate TAPNAV in MuJoCo simulation and on a Unitree G1 humanoid to examine uncertainty-bounded state estimation, low-level positioning accuracy, and integrated real-world navigation.

Simulation Results

Across four maps, TAPNAV maintains estimated trajectories close to the simulator ground truth. Contacts with different surface normals around corners and pillars provide complementary position constraints, while flat-wall contacts reduce uncertainty along the surface normal and can also reduce heading uncertainty.

Map 01: Curved Corridor

TAPNAV
Odometry-only (failure case)

Map 02: Partitioned Room

TAPNAV
Odometry-only (failure case)

Map 03: Zigzag Corridor

TAPNAV
Odometry-only (failure case)

Map 04: Room with Scattered Obstacles

TAPNAV
Odometry-only (failure case)
  • Largest position uncertainty drop
  • Largest heading uncertainty drop
  • Position/yaw uncertainty
  • Position: Estimated
  • Position: Ground truth
  • Position/yaw error
  • Contact point
  • Start
  • Goal
Map 01: Curved Corridor. Estimated and ground-truth trajectories, contact points, position and yaw uncertainty and error. Map 02: Partitioned Room. Estimated and ground-truth trajectories, contact points, position and yaw uncertainty and error. Map 03: Zigzag Corridor. Estimated and ground-truth trajectories, contact points, position and yaw uncertainty and error. Map 04: Room with Scattered Obstacles. Estimated and ground-truth trajectories, contact points, position and yaw uncertainty and error.
Download figure (PDF)
TAPNAV navigation performance in simulation. Each map shows the estimated and ground-truth trajectories, tactile contact points, and the corresponding position and yaw uncertainty and error over time.

Over 20 trials per map and method, TAPNAV achieves a 95% average success rate, compared with 77.5% for Random-touch, 38.8% for Sweep-touch, and 22.5% for Odometry-only, while producing lower position error and comparable yaw error.

Real-World Experiments

We deploy TAPNAV on a Unitree G1 with tactile end-effectors in an office hallway and a kitchen. The robot makes successive turns, navigates around the counter, and actively probes surfaces with different contact geometry while moving toward the goal.

Trial 01

TAPNAV
Odometry-only (2.74× speed)

Trial 02

TAPNAV
Odometry-only (9.52× speed)
Unitree G1 real-world navigation in an office hallway and kitchen with paths and tactile probes
TAPNAV deployed on a Unitree G1 in an office hallway and kitchen. Each environment shows deployment snapshots together with the robot path and active tactile probes in the robot's egocentric view.

Drift Recovery

Drift Recovery

Low-Level Controller Accuracy

Open-loop trajectory and heading comparison between velocity and footstep tracking policies
Simulated open-loop goal tracking with velocity- and footstep-tracking policies over 30 trials per policy: executed trajectories and actual versus commanded heading.

We compare TAPNAV's footstep controller with a velocity-tracking controller on a sequence of global planar torso targets. Heading errors accumulate under velocity tracking, while the footstep controller follows the target sequence with minimal drift.

Yaw-Noise Sensitivity

Position error, yaw error, and success rate under increasing yaw noise
Yaw-noise sensitivity: position error, yaw error, and success rate. Bands and bars show mean +/- standard error and 95% confidence intervals.

Under increasing IMU heading noise, TAPNAV retains the lowest mean position and yaw errors and the highest success rate among the evaluated methods.

Conclusion

TAPNAV integrates tactile active probing, belief-space localization, uncertainty-aware global planning, and whole-body control for vision-denied humanoid navigation. Simulation and hardware experiments demonstrate goal-directed navigation using tactile feedback, with improved navigation success and lower pose error than simulation baselines. The current system assumes a known static floor plan and a unimodal Gaussian approximation of the particle belief during planning.

BibTeX

@inproceedings{anonymous2027TAPNAV,
  title     = {TAPNAV: Humanoid Navigation through Tactile Active Perception},
  author    = {Anonymous Authors},
  booktitle = {Under Review},
  year      = {2027}
}