---
title: "Tsinghua's TactileStep Teaches Humanoid Robots to Land 48% Softer"
subtitle: "A Tsinghua team gives the Unitree G1 humanoid pressure-sensing insoles and phase-aware rewards, cutting landing impact by 48.8% and impact noise by 30.1 dB."
url: "https://alphasignal.ai/news/tsinghua-s-tactilestep-teaches-humanoid-robots-to-land-48-softer"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-09-24T03:13:38.000Z"
updated: "2026-09-27T00:06:05.522Z"
type: "Paper"
topics:
  - "Robotics"
  - "Training Infra"
subtopics:
  - "Humanoid"
  - "Distributed Training"
upvotes: 184
source_url: "https://arxiv.org/abs/2609.28959"
comments: 0
truncated: true
---

# Tsinghua's TactileStep Teaches Humanoid Robots to Land 48% Softer

> A Tsinghua team gives the Unitree G1 humanoid pressure-sensing insoles and phase-aware rewards, cutting landing impact by 48.8% and impact noise by 30.1 dB.

# TactileStep teaches humanoid robots to land softly

Parkour-style humanoid controllers can clear stairs and platforms while still driving the robot’s feet into the ground. A Tsinghua University paper accepted at CoRL 2026, [TactileStep](https://tactilestep.github.io/), addresses that problem by fitting a Unitree G1 with pressure-sensing insoles and training its locomotion policy to regulate each touchdown.

## Humanoid feet have been flying blind

Most humanoid locomotion policies rely on cameras, joint positions, motor states, and inertial measurements. Those inputs reveal where the terrain is and how the body is moving, but they do not measure how pressure spreads across the sole after contact.

TactileStep adds that missing feedback. The pressure signal lets the controller reduce landing impact and maintain a broader, better-centered support patch during stance. Lower impact could reduce mechanical stress, while quieter and more stable footfalls would help robots operate around people, delicate surfaces, stairs, and narrow footholds.

Sole sensing also creates a simulation challenge. Training occurs in a simulator, while deployment uses physical pressure sensors. A policy will transfer poorly if simulated contact values do not resemble the measurements produced by the real insole.

## Pressure becomes a policy input

The researchers built a lightweight tactile simulator that converts rigid foot-terrain contacts into a pressure array resembling the insole’s output. That array produces three compact features for the action-generating network:

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [TactileStep Project Page (CoRL 2026)](https://tactilestep.github.io/)
- [TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion (arXiv)](https://arxiv.org/abs/2609.28959)
- [TactileStep HTML paper](https://arxiv.org/html/2609.28959v1)