TAPNAV: Humanoid Navigation through Tactile Active Perception

arXiv:2610.10748 · cs.RO · Submitted 2026-10-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "TAPNAV: Humanoid Navigation through Tactile Active Perception".

Rosa: The gist Navigation in vision-denied environments is challenging for humanoid robots because proprioceptive odometry drifts and localization uncertainty accumulates rapidly,

Dev: First, who's behind it and why it matters.

Paper summary: Dev: So looking at the whole TAPNAV paper, we have these authors: Huaze Liu, Zhenyu Wu, Jaehwi Jang, Junjie Sheng, Andrew Collins, Aaron Xie, Zhaoyuan Gu, Kaijie Zhu, Ding Jiang and Kevin Cai. They put forward this integrated framework that combines belief-space planning with active tactile perception for vision-denied navigation.

Rosa: The authors summarized their contributions as creating a belief-space tactile navigation planner and an active tactile probing framework that balances route progress with localization uncertainty by exploiting opportunities for tactile correction. They also highlight the integrated nature of coupling this with reinforcement learning-based whole-body locomotion control, validated in both simulation and on real hardware.

Taro: It seems they are really pushing the idea that you don't need high-end sensors like vision to solve these localization problems if you can design a way for the robot to use physical interaction intelligently. The paper shows how tactile sensing can be used not just for simple contact detection, but as a source of active sensing information.

Dev: The implication is that we might see robots navigating complex, unmapped indoor spaces—like those offices or kitchens they tested in—more reliably when they are operating in environments where vision is unreliable or unavailable. The system relies on the robot’s own physical interaction with the world to build its map and know where it is.

Rosa: So, to put it simply, TAPNAV suggests that instead of passively drifting while moving, a humanoid can actively touch things in a way that tells it precisely where it is and helps it stay on track toward its destination. It’s about using touch as a navigation tool itself.

Conclusion: Rosa: So, TAPNAV is this new method where a robot navigates without any cameras by just touching things intelligently to figure out where it is and how to get there.

Dev: Yeah, that’s the core idea—using tactile sensing as the primary way for the robot to know its location instead of relying on vision.

Taro: What I find really interesting is how they structure that whole system, coupling a global plan with this local probing action based on information gain.

Rosa: It sounds like they’re trying to solve that big problem where robots get lost in dark or cluttered spaces where sensors fail them completely.

Dev: Exactly. They maintain a belief about its pose using odometry and IMU, but the tactile probes are what actually give that belief the necessary corrections when things get ambiguous.

Taro: So if you think of it like this, the robot doesn't just move blindly; it moves purposefully to make physical contact in a way that gives it better data about its surroundings.

Rosa: And they validated this thing both in simulation and on real hardware, which is important because those are usually very different things.

Dev: Yeah, the results showed lower estimation error than their previous methods, and the real-world tests in a hallway and a kitchen look promising for actual use outside of a controlled lab setting.

Taro: It changes how we think about autonomy because it suggests that physical interaction itself can be the primary source of navigational intelligence when other sensors are missing.

Rosa: So, basically, this paper shows that active probing with touch is a viable way to give humanoids reliable navigation cues without needing expensive visual systems.

Dev: It really puts the focus on how much information you can extract from simple physical contact versus relying on complex digital processing.

Taro: We're going to look next at exactly what kind of physical interactions they are choosing for these probes and how those choices actually affect the robot's ability to localize.

Georgia Institute of Technology

cs.RO

Submitted: 2026-10-07

Updated: 2026-10-07

Comments: 9 pages, 10 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: The gist Navigation in vision-denied environments is challenging for humanoid robots because proprioceptive odometry drifts and localization uncertainty accumulates rapidly, and TAPNAV presents a

Key concepts

Belief-Space Planning
This involves maintaining a probabilistic estimate (like a particle filter) of the robot's exact position in space. The global planner then chooses paths that are designed to reduce this uncertainty by finding areas where tactile contact can provide useful localization information.
Active Tactile Probing Framework
Instead of just moving blindly, the robot actively decides where to touch next. It calculates which probing action will give the most useful new information about its location, balancing the benefit of getting a better estimate against the cost of moving.
Expected Information Gain (EIG)
This is a metric used by the local planner to score different potential actions (probes). The action with the highest EIG is chosen because it maximizes the expected reduction in localization uncertainty, guiding the robot toward a more accurate pose estimate.
Low-Level Controller
This component manages the physical movement of the robot's body. It consists of separate policies for controlling the upper and lower body, allowing the robot to perform precise movements like tracking a target while simultaneously adjusting its foot placements for stability.

Terminology

Summary

The gist Navigation in vision-denied environments is challenging for humanoid robots because proprioceptive odometry drifts and localization uncertainty accumulates rapidly, and TAPNAV presents a tactile active-perception framework that enables humanoid navigation toward a goal by actively probing surrounding structures without relying on vision.

How it works

TAPNAV is an integrated framework that guides a humanoid to a goal using tactile sensing without relying on cameras or LiDAR The system maintains a pose belief from odometry, IMU, and tactile contact observations It couples uncertainty-aware global route planning with information-gain-driven local probing.

Belief-Space Planning and Active Localization

The system maintains a pose belief from odometry, IMU, and tactile contact observations This belief is represented by a particle filter or approximated as a Gaussian distribution whose mean is computed from weighted particles. The global planner searches for routes that keep predicted localization uncertainty bounded by exploiting opportunities for tactile correction.

Active Tactile Probing Framework

The local planner selects probe actions that maximize expected information gain (EIG) A probe moves the end-effector along a straight trajectory in a direction relative to the torso heading until contact is detected or a travel limit is reached. The expected information gain is calculated by scoring each candidate action by its expected information gain (EIG) over these two outcomes: at = arg max a∈A EIGt(a), = arg max a∈A h pt(a) IGc t(a) + (1 − pt(a)) IGf t(a).

Low-Level Controller

The low-level controller is composed of an upper and a lower-body policy, jointly trained with shared whole-body proprioception. The upper-body policy tracks the target end-effector pose and force, while the lower-body policy tracks footstep placements. This allows for frequent stop-and-go maneuvers while maintaining contact with the surroundings.

Experimental Validation

TAPNAV achieves lower state estimation error and a higher task completion rate than baselines. In simulation, TAPNAV maintains estimated trajectories closely following the ground-truth states measured from the simulator, indicating accurate localization during navigation. Real-world experiments on a Unitree G1 demonstrated successful navigation trials in both an office hallway and a kitchen environment. The results show that TAPNAV exploits active and informative tactile contacts to reduce localization uncertainty.

The study concludes that actively planning physical interactions with the environment can provide localization cues for reliable humanoid navigation without vision.

Improvements for AI systems

  1. Bold header: Belief-Space Tactile Navigation Planning Improvement

The system can be improved by relaxing the approximation of particle belief as a unimodal Gaussian during planning, which is currently stated as a current limitation. This extension would allow the framework to handle multimodal beliefs and partially known maps more accurately.

  1. Bold header: Active Probing Framework Refinement

The local planner's selection mechanism can be made more robust by incorporating the full expected information gain calculation, specifically maximizing EIGt(a) = arg max a∈A h pt(a) IGc t(a) + (1 − pt(a)) IGf t(a) for both contact and contact-free outcomes. This ensures the probe action is selected based on the true expected reduction in uncertainty, rather than relying solely on simplified models.

  1. Bold header: Global Planner Uncertainty Bounding

The global planner can be enhanced by refining how it predicts localization uncertainty along candidate routes, specifically by using a simplified observation model that accounts for contact likelihoods when a surface is within probing range. This allows the planner to more effectively constrain localization uncertainty and favor routes that exploit opportunities for tactile correction.

  1. Bold header: Low-Level Controller Generalization

The low-level controller can be improved by training the upper-body and lower-body policies with a broader set of goals, moving beyond the current specific goal definitions to potentially improve generalized whole-body coordination under more complex interaction scenarios.

Sources

Related papers