TactileStep: Sole Tactile Learning for Regulating Foot–Terrain Interaction in Humanoid Locomotion

Overview

Method

TactileStep architecture. Depth, proprioception, and sole contact features (normal force, contact area, and center of pressure) feed an actor with dense and sparse critics. Phase-conditioned rewards cover swing, pre-landing, landing, and stance. — Overview of TactileStep. Our central design is to turn sole pressure into a deployable contact-state representation for policy learning. We first construct a lightweight tactile simulator that maps rigid foot–terrain contacts to a pressure array, from which contact features are extracted and exposed to the policy during both training and deployment. Gait phases are estimated online and phase-conditioned rewards are designed to regulate touchdown impact and stance support.

Abstract

Humanoid parkour policies can traverse various terrains, but task completion may mask challenges of harsh landings, edge contacts, and unstable stance contacts. Humans naturally regulate foot–terrain interaction through tactile feedback, modulating contact compliance according to terrain stiffness. This highlights a key domain gap between humans and humanoid robots: the absence of rich tactile sensing in most humanoid systems. We address this problem with TactileStep, a deployable tactile learning framework that brings sole pressure sensing into humanoid locomotion control for softer touchdowns and more stable support. TactileStep aligns tactile simulation with the real pressure insole, allowing the policy to learn from the same contact features available on hardware. During training, we use tactile and motion cues to recognize different foot-contact phases and apply phase-aware rewards that encourage safer landing and more stable stance. Evaluated in simulation and on a Unitree G1 humanoid across diverse terrains, TactileStep reduces peak touchdown force by up to 48.8% and peak A-weighted impact noise by up to 30.1 dB over a strong perceptive baseline, while increasing stance contact area by up to 23.8%.

Locomotion across terrains

Platform ascent

Platform descent

Stair descent

Flat ground

Slope ascent

Slope descent

Key results

Softer landings, steadier support, and reliable traversal.
Evaluated in simulation and on the Real Unitree G1.

In simulation

Lower landing impact across terrains

Radar chart of simulated landing impact force in newtons across flat ground, slopes, upstairs, downstairs, and platform ascent and descent. TactileStep has lower impact force than both the variant without soft-landing rewards and the Hiking baseline on all six terrains.

In simulation

Broader, more centered support

Swipe to explore →

Simulation evaluation of support quality on terrains.
Terrain	Method	Contact area ratio ↑	CoP margin [mm] ↑
Flat	TactileStep	0.929 ± 0.054	35.79 ± 0.54
	w/o stable	0.750 ± 0.068	29.90 ± 2.34
	Baseline	0.912 ± 0.020	35.21 ± 0.14
Slope	TactileStep	0.805 ± 0.044	34.24 ± 0.37
	w/o stable	0.624 ± 0.051	27.55 ± 1.72
	Baseline	0.804 ± 0.023	34.20 ± 0.11
Upstairs	TactileStep	0.536 ± 0.042	29.05 ± 1.16
	w/o stable	0.462 ± 0.040	25.56 ± 1.35
	Baseline	0.480 ± 0.012	26.03 ± 0.40
Downstairs	TactileStep	0.545 ± 0.051	29.43 ± 1.02
	w/o stable	0.455 ± 0.048	25.91 ± 1.20
	Baseline	0.487 ± 0.017	27.07 ± 0.36

In simulation

Traversal performance and trade-offs

Swipe to explore →

Standard locomotion metrics in simulation.
Terrain	Method	Success rate	Velocity RMSE [m/s]	Traversal time [s]	Energy [J]
Stair up	TactileStep	100%	0.24 ± 0.02	5.7 ± 0.4	1126.3 ± 66.6
Stair up	Baseline	99.98%	0.21 ± 0.02	5.6 ± 0.3	802.2 ± 97.0
Stair down	TactileStep	100%	0.20 ± 0.02	6.1 ± 0.6	1232.9 ± 91.0
Stair down	Baseline	99.98%	0.19 ± 0.03	6.0 ± 0.5	887.4 ± 110.4
Platform up	TactileStep	100%	0.24 ± 0.02	4.9 ± 0.5	939.8 ± 50.5
Platform up	Baseline	99.93%	0.23 ± 0.03	4.9 ± 0.6	867.1 ± 72.8
Platform down	TactileStep	100%	0.22 ± 0.02	4.9 ± 0.6	1033.0 ± 73.6
Platform down	Baseline	92.26%	0.21 ± 0.15	4.9 ± 0.8	1005.3 ± 119.7
Flat	TactileStep	99.98%	0.20 ± 0.08	6.5 ± 3.3	834.4 ± 165.4
Flat	Baseline	99.19%	0.17 ± 0.03	5.6 ± 2.2	644.5 ± 76.1
Slope	TactileStep	99.63%	0.21 ± 0.17	6.0 ± 2.8	995.7 ± 126.1
Slope	Baseline	99.63%	0.19 ± 0.05	5.2 ± 2.1	786.3 ± 47.9

On the real robot

Softer, quieter contact on hardware

Swipe to explore →

Real-world results across diverse terrains.
Terrain	Method	Impact force [N] ↓	Peak A-weighted noise [dB] ↓	Contact area ratio ↑
Stair up	Baseline	499.9 ± 40.7	101.2 ± 3.7	0.470 ± 0.013
	w/o tac. obs.	375.2 ± 47.3	85.1 ± 3.4	0.421 ± 0.028
	TactileStep	349.8 ± 32.5	72.2 ± 2.8	0.482 ± 0.016
Stair down	Baseline	359.3 ± 23.2	97.2 ± 1.2	0.483 ± 0.020
	w/o tac. obs.	353.3 ± 38.6	84.9 ± 3.9	0.417 ± 0.044
	TactileStep	344.1 ± 20.1	67.1 ± 1.4	0.598 ± 0.027
Platform up	Baseline	695.0 ± 49.4	110.1 ± 6.0	–
	w/o tac. obs.	371.5 ± 50.5	98.2 ± 5.3	–
	TactileStep	355.7 ± 39.8	83.0 ± 4.1	–
Platform down	Baseline	608.9 ± 38.6	90.8 ± 5.2	–
	w/o tac. obs.	445.7 ± 54.1	87.5 ± 4.4	–
	TactileStep	404.6 ± 43.2	75.9 ± 3.6	–
Flat	Baseline	201.3 ± 23.3	90.6 ± 2.5	0.453 ± 0.011
	w/o tac. obs.	202.1 ± 32.4	85.9 ± 2.9	0.362 ± 0.035
	TactileStep	191.3 ± 17.7	66.5 ± 1.6	0.510 ± 0.014
Slope	Baseline	255.0 ± 34.9	83.4 ± 1.0	0.486 ± 0.026
	w/o tac. obs.	204.7 ± 42.1	77.1 ± 2.0	0.490 ± 0.033
	TactileStep	202.0 ± 31.0	69.9 ± 1.1	0.499 ± 0.030

BibTeX

@misc{wang2026tactilestepsoletactilelearning,
       title={TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion},
       author={Zizhuo Wang and Ming-ju Lee and Shaoting Zhu and Haozhe Lou and Hang Zhao and Yiming Li},
       year={2026},
       eprint={2609.28959},
       archivePrefix={arXiv},
       primaryClass={cs.RO},
       url={https://arxiv.org/abs/2609.28959},
}