Source-linked AI summary
TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion
Zizhuo Wang, Ming-ju Lee, Shaoting Zhu, Haozhe Lou, Hang Zhao, Yiming Li
TL;DR
Humanoid parkour policies can complete traversal while leaving harsh landings, edge contacts, and unstable stance contacts unresolved. TactileStep gives the policy deployable sole-pressure features and applies phase-aware rewards; experiments report reduced touchdown impact and noise and improved stance support across diverse terrains.
Problem
Traversal success can mask harsh landings, risky edge contacts, and unstable stance, while deployed policies lack direct access to the sole-pressure state relevant to contact quality.
Source: §1 · p. 2Method
TactileStep aligns simulated and hardware sole-pressure features for policy observation and uses phase-conditioned rewards to regulate touchdown impact and stance support.
Source: §1 · p. 2Results
Takeaways & Limitations
Sole tactile feedback makes foot–terrain contact quality part of closed-loop humanoid parkour control, beyond traversal success alone.
Source: §5 · p. 9Generalization to substantially faster motions remains unexplored because training and evaluation use a bounded command range.
Source: §6 · p. 9
Abstract
from arXiv · showhide
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%.
1 Introduction
TactileStep brings hardware-available sole-pressure features into humanoid parkour control, addressing contact-quality problems that terrain traversal success alone can conceal. Its phase-aware rewards target touchdown impact and stance support, and the framework is validated on a Unitree G1 across different terrains.
1 Introduction
- Harsh touchdown, edge-biased footholds, and small-area or offset-CoP stance contacts can undermine support and propagate into whole-body instability.These local contact problems can persist even when a parkour task is completed.Source: §1 · p. 2
- Sole-pressure features—normal force, contact area, and CoP—give the policy a hardware-available view of contact quality after touchdown.This representation addresses the mismatch between pre-contact terrain sensing and the sole-pressure state needed for landing and support.Source: §1 · p. 2
- Phase-conditioned rewards regulate downward foot motion before contact, impact at touchdown, and support stability during stance.The step is divided into SWING, PRE-LANDING, LANDING, and STANCE phases.Source: §1 · p. 2
- TactileStep validation on a Unitree G1 shows lower touchdown impact and improved tactile support metrics on different terrains.The framework makes foot–terrain contact quality an objective alongside traversal.Source: §1 · p. 2
2 Related Work
Related work addresses foothold feasibility through terrain geometry, tactile sensing through representations of varying complexity, and soft landing through hardware or learning-based control. TactileStep combines sole-pressure features with phase-conditioned rewards to regulate both landing and stance support.
2 Related Work
- TactileStep’s overview links rigid foot–terrain contacts to a pressure array, policy-facing contact features, online gait phases, and phase-conditioned rewards.The features are used in both training and deployment to regulate touchdown impact and stance support.Source: Figure 2 · p. 3
- Geometric foothold objectives reduce risky edge contacts, but feasible placement alone does not ensure stable post-touchdown support.Sole-pressure features are used to regulate both foothold acquisition and stance support [8] [9] [10] [20] [22] [25].Source: §2 · p. 3
- Tactile simulation trades representational richness against scalability by distributing rigid-body forces over sole taxels and learning on normal force, contact area, and CoP features.Both policy learning and sim-to-real alignment use these features rather than the full plantar-pressure distribution.Source: §2 · p. 3
- Prior soft-landing methods include compliant hardware, local contact control, and learning-based approaches, while QuietWalk is limited to regular walking and leaves deployment-time sole feedback unused [13, 14, 32–37].TactileStep exposes contact features to the policy and applies phase-conditioned rewards on parkour terrains such as stairs and high platforms.Source: §2 · p. 3
3 Method
TactileStep combines deployable sole-pressure observations with phase-conditioned rewards to regulate foot–terrain interaction. Its method pairs a partially observable locomotion policy with tactile simulation and phase-aware contact objectives.
3 Method
- The actor receives deployable tactile features alongside proprioceptive and depth histories, while simulator-specific contact signals remain privileged to critics and reward computation.Source: §3 · p. 3, §3.1 · p. 4
3.1 Problem Formulation
- The policy uses PPO in a POMDP, outputs target joint positions for PD control, and trains with two critics that separate dense from sparse reward signals.Reward terms cover task, regularization, safety, and motion-prior objectives.Source: §3.1 · p. 4
3.2 Tactile Simulation
- The tactile simulator estimates taxel–terrain gaps and normals to distribute foot forces, then diffuses taxel loads to smooth sparse or unstable rigid-body contacts.Source: §3.2 · pp. 4–5
- From the smoothed pressure map, the method extracts normal force, contact-area ratio, and CoP, compact features described as less sensor-specific for sim-to-real transfer.Figure 3 depicts the raycast-derived pressure map and its contact-area and CoP outputs.Source: Figure 3 · p. 4, §3.2 · p. 5
- During stair descent under the tactile-independent Baseline, simulated and measured normal-force, contact-area, and CoP-margin traces have similar ranges and trends despite residual differences.Source: §3.2 · p. 5, Figure 4 · p. 5
3.3 Phase-Aware Rewards for Soft Landing and Support Stability
- Each foot is assigned SWING, PRE-LANDING, LANDING, or STANCE using contact and motion cues, with the transition phases capturing windows around touchdown.Source: §3.3 · p. 6
- Phase estimates route tactile rewards to relevant gait segments, preventing touchdown penalties during swing and stance-stability terms before contact.Source: §3.3 · p. 6
- PRE-LANDING and LANDING rewards discourage aggressive approach speed, impact force, and rapid or peak force increases; STANCE rewards favor broad, centered, temporally stable support.The stance terms discourage partial edge contacts and unstable pressure migration.Source: §3.3 · p. 6
4 Experimental Results
TactileStep improves touchdown and stance contact quality in simulation and on hardware, while standard locomotion metrics remain comparable to the baseline. The experiments also test which tactile observations and reward components contribute to those results.
4.1 Experimental Setup
- The evaluation compares TactileStep with the Hiking vision-based baseline and ablations removing tactile observations, soft-landing rewards, or stable-support rewards, using impact and stance-support metrics.Landing impact force is measured at touchdown; contact area ratio measures sole support during stance.Source: §4.1 · p. 7
4.2 Simulation Results
- In simulation, TactileStep has the lowest landing impact force across tested terrains; contact area and CoP margin improve over the no-stable-support variant and generally over baseline.Support gains are larger on stairs than on flat ground and slopes; removing soft-landing rewards consistently increases impact force.Source: §4.2 · p. 8, Figure 6 · p. 8, Table 1 · p. 8
- Simulation success rates are at least as high as Baseline, while velocity-tracking error and traversal time are generally slightly higher but comparable; energy consumption is higher.The paper suggests longer traversal times and more active joint regulation may contribute to the higher energy use.Source: §4.2 · p. 8, Table 2 · p. 8
4.3 Real-World Results
- On hardware, TactileStep produces lower landing impact force and peak A-weighted noise than Hiking, with broader stance contact area across tested terrains.The force and acoustic results indicate softer touchdowns; the larger contact area indicates the improvement does not come from weakening contact.Source: §4.3 · p. 9, Table 3 · p. 9
- The hardware ablation remains weaker than TactileStep in key measurements, especially acoustic noise and contact area, when online tactile observations are removed.The variant often still improves over Hiking.Source: §4.3 · p. 9
5 Conclusion
TactileStep treats foot–terrain interaction as a central locomotion objective, using deployable pressure-based features and phase-aware rewards. Simulation and Unitree G1 experiments show reduced impact and noise and improved support across diverse terrains.
5 Conclusion
- TactileStep combines hardware-aligned pressure features with phase-aware rewards to regulate contact quality from pre-landing through stance, improving touchdown impact, acoustic noise, and stance support across diverse terrains.Experiments were conducted in simulation and on a Unitree G1 humanoid.Source: §5 · p. 9
6 Limitations and Future Work
The evaluation and model assumptions leave several deployment boundaries unresolved, including motion speed, terrain material, sensor durability, and recovery behavior.
6 Limitations and Future Work
- Training and evaluation use a bounded command range, leaving generalization to substantially faster motions with shorter contact duration and larger impact transients unexplored.The phase estimator also relies on hand-designed contact and motion cues; learned phase representations may improve robustness across speeds and terrains.Source: §6 · p. 9
- The tactile simulator assumes rigid foot–terrain contact, and its fidelity on deformable or granular surfaces has not been validated.Source: §6 · p. 9
- The study does not evaluate long-term pressure-insole durability, drift, or recalibration, and does not systematically characterize failures or recovery behavior.Source: §6 · p. 9
A.1 Evaluation Protocol in Simulation
Simulation policies are tested in parallel across six terrain categories under matched terrain instances. Platform transitions are assessed for landing impact rather than sustained stance support.
A.1 Evaluation Protocol in Simulation
- Each evaluation run uses a 5-by-10 terrain grid and one episode from each of 4096 parallel environments; reported metrics are means over those episodes.Source: §A.1 · p. 13
- Each policy is evaluated separately on flat ground, slopes, upstairs, downstairs, up-platforms, and down-platforms, with one terrain category per run.Figure 7 depicts the parallel evaluation scenes.Source: §A.1 · p. 13, Figure 7 · p. 13
- Policies use cached terrain generation and a fixed random seed within each terrain category, so comparisons use identical terrain instances.Source: §A.1 · p. 13
- Contact-area ratio and CoP margin are not reported for platform step-up or drop-down because those single-step trials target LANDING, not sustained STANCE support.Source: §A.1 · p. 13
A.2 Single-Critic Ablation
The dual-critic policy reduces touchdown force and improves stance-support metrics versus a single-critic variant on most terrains. The ablation compares these variants under the same reward terms and training budget.
A.2 Single-Critic Ablation
- The dual-critic policy consistently reduces touchdown force and improves stance-support metrics on most terrains compared with the single-critic variant.The comparison uses the same reward terms and training budget.Source: §A.2 · p. 13, Table 4 · p. 14
- Separating dense and sparse reward returns is suggested to help learn contact-quality objectives without sacrificing locomotion stability.This is the authors’ interpretation of the dual-critic comparison.Source: §A.2 · p. 13
- Table 4 compares the dual-critic policy with a single-critic variant, with lower landing impact force and higher CoP margin and contact-area ratio marked as better.Source: Table 4 · p. 14
B Training Details
Reward design separates dense, continuous learning signals from sparse, event- or constraint-linked signals, assigning each group a critic while mixing their advantages for actor updates. A terrain curriculum promotes or demotes environments by episode-level velocity-tracking performance as difficulty increases.
B.1 Reward Design
- The reward manager computes each reward group from weighted raw reward terms, and an AMP auxiliary reward regularizes motion alongside the environment rewards.The weights multiply raw reward terms; AMP is optimized with dense and sparse environment rewards.Source: §B.1 · pp. 14–15
- Separate critics estimate dense and sparse reward returns, and their advantages are mixed for actor updates to reduce value-estimation interference between continuous shaping and sparse contact, task-completion, or safety objectives.Group membership reflects the temporal learning signal, not task semantics or phase-conditioning.Source: §B.1 · p. 14
- Dense rewards provide frequent, continuous signals such as command tracking, posture regulation, smoothness penalties, and tactile shaping, while sparse rewards become informative at events, gates, or constraint violations.Phase-conditioned rewards can still be dense when active at every step of their phase.Source: §B.1 · p. 14
B.2 Terrain Curriculum
- The terrain curriculum initializes a fixed grid of 10 difficulty rows and 20 terrain columns, then updates environment terrain origins while keeping geometry fixed.Rows encode difficulty levels and columns encode terrain categories.Source: §B.2 · p. 15
- Environments start at low-to-medium difficulty and are promoted or demoted according to episode-level velocity-tracking performance, gradually exposing the policy to stronger disturbances and more demanding support.Rough-flat terrains target stance robustness, while stairs and platforms target swing clearance, touchdown regulation, and support quality across height changes.Source: §B.2 · p. 15
B.3 Training Hyperparameters
Training uses a shared PPO budget and dual-critic AMP-PPO, with domain randomization and tactile-noise perturbations intended to expose the policy to sensing and physical uncertainty. The actor receives compact tactile features, and higher-rate insole tests give impact-force measurements close to those from the standard wireless protocol.
B.3 Training Hyperparameters
- All policies use 2048 parallel environments for 50,000 iterations, collecting 24 steps per environment and 49,152 transitions per PPO update unless otherwise specified.Baselines and ablations share the same PPO budget and optimization settings unless otherwise stated.Source: §B.3 · p. 15, Table 8 · p. 17
- Dual-critic AMP-PPO assigns dense and sparse reward returns to separate critic heads, mixes their advantages for actor updates, and optimizes AMP as an additional motion-regularization objective.The dense and sparse reward groups provide the respective critic targets.Source: §B.3 · pp. 15–16, Table 5 · p. 15, Table 6 · p. 16
- The terrain-curriculum parameter is interpolated from the easiest to the hardest row.Source: Table 7 · p. 16
B.4 Domain Randomization
- Domain randomization covers physical parameters, actuation, initial states, disturbances, observations, perception, tactile sensing, and command generation; actor inputs are corrupted while critic observations remain clean and privileged.The observation treatment is used to stabilize value estimation while exposing the deployed policy to sensing uncertainty.Source: §B.4 · p. 16, Table 9 · p. 17
C.1 Tactile Noise Modeling in Simulation
- Training perturbs taxel force maps for force-scale uncertainty, taxel-coordinate calibration error, and short measurement delay before extracting compact tactile features.Coordinate perturbations stay fixed within an episode, while delay is sampled per environment and foot; noisy forces are not renormalized to the clean total.Source: §C.1 · p. 16, Table 10 · p. 18
C.2 Feature Normalization
- The actor uses compact per-foot tactile features rather than the full taxel map, including contact-area ratio, normalized normal force, and normalized CoP.Contact area is clipped to [0, 1], and CoP is clipped to [−1, 1] and set to zero below the stated low-force threshold.Source: §C.2 · pp. 16–18
- A two-frame tactile history contributes 16 actor-observation dimensions, while the training critic additionally receives normalized vertical foot velocity that the deployed actor omits.Source: §C.2 · p. 18
C.3 Insole Sampling Rate and Peak-Force Validation
- At 100 Hz, measured impact forces were 199.2 ± 14.9 N on flat ground and 427.7 ± 20.1 N on platform drop-down, compared with 191.3±17.7 N and 404.6±43.2 N under the 25-Hz protocol.The authors report that the agreement supports consistent rising-edge estimates; all policy comparisons use the same wireless sensing and processing pipeline.Source: §C.3 · p. 18