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

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%.

• The paper introduces TactileStep, a learning-based approach that equips a Unitree G1 humanoid with thin plantar pressure insoles to regulate foot-sole pressure and optimize foot-terrain interaction across various gait phases.
• TactileStep reduces impact force by up to 48.8% and peak noise by up to 30.1 dB and improves stance contact area by up to 23.8%.
• The research shows that tactile feedback is crucial for reliable regulation of the contact state during humanoid locomotion, particularly on variable and challenging terrains.

Problem formulation and motivation

“TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion” (2609.28959) addresses a specific deficiency in current perceptive humanoid locomotion: successful traversal does not imply physically desirable foot–terrain interaction. A vision-based policy may identify a feasible foothold and complete a stair or platform maneuver while still producing excessive touchdown forces, edge-dominated contacts, or unstable stance support. These effects are consequential for whole-body stability, actuator loading, structural durability, acoustic emissions, and operation near humans.

The paper’s central claim is that contact quality should be treated as a closed-loop control objective rather than as an indirect consequence of geometric foothold selection. Vision and depth sensing provide pre-contact information about terrain geometry, but they do not directly reveal the realized pressure distribution after touchdown. Proprioceptive measurements and estimated ground-reaction forces provide related signals, yet do not expose the spatial distribution of load across the sole. TactileStep therefore equips a Unitree G1 humanoid with thin plantar pressure insoles and provides the policy with deployable features describing the realized contact state.

The framework is organized around four phases for each foot: Swing, Pre-Landing, Landing, and Stance. This phase decomposition is important because the relevant control objectives differ temporally. During Pre-Landing, the controller should reduce excessive downward velocity and acceleration. During Landing, it should suppress impact force and force-rate transients. During Stance, it should establish broad support, maintain a centered center of pressure (CoP), and avoid abrupt pressure migration. The method consequently rejects a single undifferentiated contact penalty in favor of phase-conditioned reward terms.

Figure 1: TactileStep maps sole pressure to deployable contact features and routes phase-conditioned tactile objectives through the locomotion policy.

Deployable tactile representation

A major design decision is to avoid exposing the actor to simulator-specific contact quantities that are unavailable on hardware. Instead, TactileStep constructs a lightweight tactile simulator whose output is processed into the same low-dimensional features measured by the real insole. Each simulated sole contains 60 taxels. Raycasts estimate the distance and normal alignment between each taxel and the terrain, and the resultant rigid-body normal contact force is distributed across taxels according to proximity and orientation. A spatial diffusion step then spreads load over neighboring taxels to approximate the compliant load distribution of a real sole and to suppress isolated point contacts generated by rigid-body collision solvers.

The policy does not receive the full pressure map. It receives, for each foot, normalized normal force, contact-area ratio, and two-dimensional CoP. These features are physically interpretable and less sensitive than raw taxel readings to sensor layout, manufacturing variation, calibration, and sim-to-real mismatch. A two-frame tactile history is concatenated with proprioceptive history and depth-observation history. The actor therefore receives contact feedback while retaining the perceptual information required for foothold selection and terrain traversal.

This compression imposes an explicit information bottleneck. The representation discards fine-grained pressure topology, shear force, and other potentially useful tactile variables. The paper argues that force magnitude, occupied contact area, and CoP contain the principal information needed for touchdown regulation and stance support, but it does not establish that these features are sufficient for arbitrary contact geometries or deformable surfaces.

The simulator is deliberately not a soft-body model. It approximates compliant pressure formation through force weighting and diffusion, which permits massively parallel reinforcement learning at manageable computational cost. Feature-level comparisons under a tactile-independent baseline show similar ranges and temporal trends between simulation and hardware during stair descent, although residual discrepancies remain. This alignment strategy is therefore pragmatic rather than physically complete: it attempts to match the policy-relevant statistics instead of reproducing the full mechanics of plantar deformation.

Phase-aware reinforcement learning

TactileStep formulates locomotion as a POMDP and trains a PPO policy whose 29-dimensional action is interpreted as target joint positions for a PD controller. The total objective combines task tracking, regularization, safety, AMP motion-prior terms, and tactile contact-quality objectives. Training uses 2,048 simulated humanoids and 50,000 PPO iterations, with domain randomization applied to dynamics, actuation, observations, depth sensing, tactile force scale, taxel positions, and measurement delay.

The gait-phase estimator combines tactile force, contact-area activation, foot height, and vertical foot velocity. A foot without contact is classified as either Swing or Pre-Landing depending on whether it is descending toward the terrain. Once contact is detected, a fixed landing window distinguishes Landing from Stance. This estimator is hand-designed and depends on thresholds, but it gives the reward system a physically meaningful temporal structure without requiring an externally imposed periodic gait clock.

The Pre-Landing objective penalizes excessive downward velocity and acceleration. The Landing objective penalizes both the instantaneous normalized force and its positive increment, with additional peak statistics over the landing window. The Stance objective rewards contact area and CoP margin while penalizing abrupt CoP displacement. These terms encode a specific landing–support trade-off: reducing impact should not be achieved by producing a weak or poorly centered contact.

The method also uses dual critics. Dense rewards, such as velocity tracking, posture regulation, action smoothness, and continuously active tactile shaping, are assigned to one critic; sparse or event-triggered signals, including peak landing penalties, target completion, safety violations, and contact events, are assigned to another. Their advantages are mixed for the actor update. The rationale is to reduce value-estimation interference between continuous locomotion objectives and intermittent contact-quality events. The appendix reports that the dual-critic policy outperforms a single-critic variant in most contact metrics. For example, on stair descent, touchdown force decreases from 579.83 N with a single critic to 320.74 N with the dual critic, while contact area increases from 0.516 to 0.545. This result supports the architectural motivation, although it does not isolate whether the benefit arises specifically from critic separation or from associated optimization effects.

Simulation evaluation

Simulation experiments use 4,096 evaluation trials per policy and terrain, with six terrain categories: flat ground, slopes, stair ascent, stair descent, platform step-up, and platform drop-down.

Figure 2: Parallel simulation evaluation across flat, sloped, stair, and platform terrains.

TactileStep consistently improves contact quality relative to both the external perceptive baseline and the relevant ablations. For sustained support on flat ground, the contact-area ratio reaches 0.929, compared with 0.912 for the baseline and 0.750 without the stable-support reward. On slopes, the corresponding values are 0.805, 0.804, and 0.624. The CoP margin follows the same pattern: TactileStep achieves 35.79 mm on flat ground and 34.24 mm on slopes, while the no-stable-support variant reaches only 29.90 mm and 27.55 mm, respectively.

The gains are more pronounced on stairs, where foothold geometry makes post-contact support more difficult. During stair ascent, TactileStep obtains a contact-area ratio of 0.536, compared with 0.480 for the baseline and 0.462 without stable-support rewards. Its CoP margin is 29.05 mm, compared with 26.03 mm for the baseline and 25.56 mm for the ablation. During stair descent, the contact-area ratio is 0.545 versus 0.487 for the baseline, while the CoP margin is 29.43 mm versus 27.07 mm. These results indicate that the support reward has its greatest effect where geometric discontinuities make complete and centered sole contact difficult.

The impact-force ablation attributes touchdown improvement primarily to the soft-landing reward. Removing the soft-landing terms consistently increases simulated impact force across terrain types. Removing tactile observations also degrades the hardware-relevant contact metrics, demonstrating that the reward alone is insufficient if the deployed actor cannot observe the realized contact state.

The locomotion objective is not substantially sacrificed. Success rates are approximately 100% for TactileStep on all reported simulated terrain categories: 100% on stair ascent, stair descent, platform step-up, and platform drop-down, 99.98% on flat ground, and 99.63% on slopes. The baseline is comparable on most terrains but falls to 92.26% on platform drop-down. Velocity-tracking error and traversal time are generally slightly worse for TactileStep. The clearest cost is energy: on stair ascent, energy increases from 802.2 J for the baseline to 1,126.3 J for TactileStep; on stair descent, it increases from 887.4 J to 1,232.9 J. Thus, the method improves contact quality without reducing task success, but it does so with a nontrivial energetic penalty, plausibly caused by more active regulation and marginally longer or more corrective motions.

Hardware deployment

The hardware experiments deploy the learned controllers on a Unitree G1 with pressure insoles. Each condition uses 20 samples. Impact force is measured from the first rising edge after touchdown, while acoustic measurements use peak A-weighted sound pressure level. The insole operates wirelessly at 25 Hz; supplementary 100 Hz tests on flat ground and platform drop-down produce comparable force estimates, providing some validation of the measurement protocol.

Figure 3: Hardware touchdown-force extraction and acoustic measurement configuration during stair locomotion.

The strongest hardware evidence concerns impact reduction. On stair ascent, TactileStep reduces impact force from 499.9 N for the perceptive baseline to 349.8 N, a reduction of approximately 30.0%. On platform ascent, the reduction is from 695.0 N to 355.7 N, approximately 48.8%, which is the largest reported force improvement. On platform descent, force decreases from 608.9 N to 404.6 N, approximately 33.6%. The corresponding acoustic reductions are also substantial:

Terrain	Baseline impact	TactileStep impact	Baseline noise	TactileStep noise
Stair ascent	499.9 N	349.8 N	101.2 dB	72.2 dB
Stair descent	359.3 N	344.1 N	97.2 dB	67.1 dB
Platform ascent	695.0 N	355.7 N	110.1 dB	83.0 dB
Platform descent	608.9 N	404.6 N	90.8 dB	75.9 dB
Flat	201.3 N	191.3 N	90.6 dB	66.5 dB
Slope	255.0 N	202.0 N	83.4 dB	69.9 dB

The maximum acoustic improvement is 30.1 dB on stair descent. Because the background noise floor is approximately 55–60 dB and is dominated by the robot’s fan, these measurements concern impact-induced peaks rather than an isolated measurement of foot noise. Nevertheless, the consistent reduction in force and sound supports the claim that TactileStep produces physically softer touchdowns rather than merely changing the acoustic signature.

The support results show that lower impact does not require reduced load-bearing contact. On stair ascent, contact area increases from 0.470 for the baseline to 0.482; on stair descent, it increases from 0.483 to 0.598, a relative improvement of approximately 23.8%. On flat ground, the increase is from 0.453 to 0.510. Slope performance is similar but still favorable, increasing from 0.486 to 0.499. Contact-area measurements are not reported for platform step-up and drop-down because those trials emphasize a single transition rather than sustained stance.

The tactile-observation ablation is informative. On platform ascent, removing online tactile observations yields 371.5 N, compared with 355.7 N for the complete method; on stair descent, it yields 353.3 N versus 344.1 N. The differences in force are sometimes modest, but the acoustic and support gaps are more pronounced. On stair descent, the no-tactile-observation policy produces 84.9 dB and a contact-area ratio of 0.417, whereas TactileStep produces 67.1 dB and 0.598. This supports the paper’s stronger claim that deployment-time tactile feedback is necessary for reliable regulation of the contact state, not merely useful as a training-time proxy.

Limitations and open questions

The evaluation is broad in terrain type but bounded in operating regime. Commands are sampled around a forward speed of 0.45–0.55 m/s with variable yaw rate, so the method’s behavior at substantially higher speeds is unresolved. Faster locomotion would shorten the Landing phase and amplify the consequences of force-estimation latency, pressure-sensor sampling, and phase-classification errors.

The tactile simulator also assumes rigid terrain and approximates sole compliance through diffusion. Its validity on deformable, granular, slippery, or highly irregular surfaces is not established. The feature representation excludes shear forces and does not model the mechanical coupling between sole deformation and pressure redistribution. Consequently, the reported sim-to-real agreement should not be interpreted as validation of the simulator’s contact mechanics beyond the tested regime.

The gait-phase estimator depends on hand-designed thresholds for force, area, height, and vertical motion. This introduces a potential failure mode under unusual contact sequences, partial footholds, rapid reversals, or simultaneous multi-contact events. The experiments do not systematically report phase-estimation errors, policy failures, recovery behavior, or the distribution of unsuccessful trials beyond aggregate success rates.

Finally, long-term pressure-insole drift, wireless reliability, sensor durability, and recalibration are not evaluated. The hardware results establish short-horizon deployment feasibility, but not robustness over extended operation. The paper therefore leaves open whether the compact features remain calibrated under wear, temperature variation, repeated high-load impacts, or changes in footwear and sole mechanics.

Conclusion

TactileStep reframes humanoid locomotion contact as a controllable state rather than a passive outcome of visual foothold selection. Its principal technical contribution is the alignment of simulated and hardware plantar sensing through compact force, contact-area, and CoP features, combined with phase-aware rewards and a dual-critic PPO architecture. Across simulation and Unitree G1 experiments, the method preserves near-perfect traversal success while reducing hardware touchdown force by up to 48.8%, peak acoustic noise by up to 30.1 dB, and increasing stance contact area by up to 23.8%. The results establish that sole tactile feedback can improve both impact mitigation and post-touchdown support, although the gains entail higher simulated energy consumption and remain bounded by assumptions about terrain rigidity, sensing calibration, gait-phase estimation, and operating speed.