Tread: A Visuo-Inertial Framework for Terrain and Locomotion Context-Aware Gait-Phase and Speed Estimation in Wearable Robotics

Debadrata Sarkar; Akash Bachhar; Aman Arora

cs.ROOct 13, 2026

Tread: A Visuo-Inertial Framework for Terrain and Locomotion Context-Aware Gait-Phase and Speed Estimation in Wearable Robotics

Authors: Debadrata Sarkar, Akash Bachhar, Aman Arora

Organizations: Human Centred Robotics and Cybernetics Group, CSIR-Central Mechanical Engineering Research Institute, Durgapur 713209, India

Abstract

A wearable assistive robot must continuously estimate terrain, locomotion mode, gait phase, and walking speed. These quantities are commonly addressed by separate, terrain-specific models, with datasets rarely collected under a common protocol. This paper presents a visuo-inertial framework that connects these estimates through a causal pipeline using a chest-mounted camera and four lower-limb inertial measurement units (IMUs). Anticipatory Terrain Perception (ATP) classifies incoming terrain with 96.3% accuracy and anticipates terrain transitions by 1.46 s on average. Terrain-Aware Locomotion Mode Identification (TALMI) achieves 93.7% and 98.8% accuracy for its level-ground and stairs experts respectively, with mean transition latencies being below 200 ms. The relevant TALMI posterior is retained as a Probabilistic Locomotion Context (PLC) vector and provided to a shared Gait Kinematic Encoder (GKE), which supports 50-Hz bilateral gait-phase estimation and 2-Hz walking-speed estimation. Leave-one-subject-out evaluation yields gait-phase RMSEs of 1.75%, 2.97%, and 4.55% for level walking, stair ascent, and stair descent respectively while estimating speed over 3.5-5.5 km/h with a RMSE of 0.218 km/h. All inertial sensing based models use sensor-fused Euler angles as input rather than raw accelerometer and gyroscope signals, reducing dependence on sensor mounting variation.

Figures & tables

Fig. 1: Wearable visuo-inertial platform: front and rear views of the instrumented participant.

Work (year)	Sensing	Inertial signals	Terrain	Locomotion modes	Gait phase	Speed	Context coupling	Indicative result
Laschowski et al. [ 2 ] (2022)	RGB vision	–	12 classes	–	–	–	Perception only	Best 12-class test accuracy: 73.2%
Tricomi et al. [ 3 ] (2023)	RGB vision + 2 IMU	acc, gyro	LG, SA, SD	–	Continuous	–	Deterministic switch	Per-class accuracy $>$ 85%
Tricomi et al. [ 4 ] (2025)	RGB vision (geometric)	acc, gyro	LG, stairs	–	–	–	Geometric context	Overall terrain-detection accuracy: 93.0 $\pm$ 1.1%
Qian et al. [ 7 ] (2022)	Depth vision + IMU	acc, gyro	LG, SA, SD	LG, SA, SD	Continuous	–	Predictive switch	Mode accuracy: $>$ 98%
Ma et al. [ 6 ] (2025)	IMU only (6)	acc, gyro	–	5 steady + transitions	–	–	None (majority filter)	Steady-mode accuracy: 99.58%
Kang et al. [ 8 ] (2021)	IMU + hip encoders	acc, gyro, joint angle	–	–	Continuous	–	None (mode-agnostic)	Gait Phase RMSE: 5.04 $\pm$ 0.79%

TABLE I: COMPARISON WITH REPRESENTATIVE WORKS

Fig. 2: (a) ATP dataset construction from ExoNet and custom frames using the terrain dominance criterion and three terrain classes. (b) Power and data architecture of the onboard edge-inference backpack.

Fig. 3: ATP pipeline using a frozen ResNet-18 feature extractor and trainable three-layer classifier head.

Class	Precision	Recall	F1-Score
LG	0.945	0.953	0.949
US	0.967	0.973	0.970
DS	0.978	0.966	0.972
Accuracy	0.963

TABLE II: ATP CLASSIFICATION REPORT ON THE HELD-OUT TEST SET

Fig. 4: ATP performance on the held-out test set. (a) Confusion matrix. (b) Accuracy within 200 ms of terrain transitions.

Fig. 5: Synchronized heel-contact and IMU traces used for annotation. (a) Level-ground sequence. (b) stairs sequence. (c,d) TALMI confusion matrices. FSR based heel-contact signals are used only for annotation.

Fig. 6: TALMI pipeline from IMU acquisition and feature extraction to temporal stacking and the LGLI and SLI softmax layers.

Expert	Class	Precision	Recall	F1 Score
LGLI	ST	0.941	0.956	0.949
	FG	0.985	0.966	0.976
	LW	0.954	0.911	0.932
	RT	0.772	0.908	0.835
	LT	0.797	0.907	0.848
	Accuracy	0.937

TABLE III: TALMI CLASSIFICATION REPORTS FOR THE LGLI AND SLI EXPERTS

Fig. 7: Coupled ATP–TALMI inference for a participant excluded from training, validation, and test partitions. (a) Mixed terrain route; (b) level-ground segment containing LGLI classes.

Fig. 8: Subject-independent TALMI evaluation. (a) Per-class recall under leave-one-subject-out cross-validation. (b) Accuracy within 800 ms of mode transitions on unseen participants.

Fig. 9: CAGPE and CAWSE pipeline with PLC formation, the shared Gait Kinematic Encoder, and the two decoder heads with their inference rates.

Fig. 10: CAGPE performance. (a)–(c) Estimated versus true gait phase for LW, SA, and SD. (d)–(e) Continuous bilateral phase estimation across stairs-to-level-walking transitions (stair ascent and stair descent, respectively).

Fig. 11: LOSO evaluation of the context-aware estimators. (a) CAGPE gait-phase RMSE for level walking, stair ascent, and stair descent; (b) CAWSE predicted versus actual walking speed over 3.5–5.5 km/h

Stage	Targeted Rate	Measured Latency (ms)
ATP	30 Hz	13.57 $\pm$ 1.22
TALMI	10 Hz	14.52 $\pm$ 0.58
CAGPE	50 Hz	15.9 $\pm$ 0.8
CAWSE	2 Hz	256.05 $\pm$ 5.35

TABLE IV: TARGETED INFERENCE RATES AND MEASURED ONBOARD LATENCY

Explore similar work

User-Tailored Learning to Forecast Walking Modes for Exosuits
Mar 16, 2026Gabriele Abbate, Enrica Tricomi, Nathalie Gierden +3Assistive Robotics Wearable Sensing
Dalle Molle Institute for Artificial Intelligence (IDSIA), USI-SUPSI, Lugano, Switzerland · Munich Institute for Robotics and Machine Intelligence (MIRMI), Department of Computer Engineering, School of Computation, Information and Technology, Technical University of Munich (TUM), Munich, Germany.
GAIT: Legged Robot Proprioceptive State Estimation with Attention over Inertial-Leg Tokens
Jun 12, 2026Young-Rang Seo, Hajun Kim, Sangmin Kim +2Legged Robotics Robot State Estimation
Korea Advanced Institute of Science and Technology, Yuseong-gu, Daejeon 34141, Republic of Korea.
PRIOR: Perceptive Learning for Humanoid Locomotion with Reference Gait Priors
Mar 19, 2026Chenxi Han, Shilu He, Yi Cheng +2Humanoid Robot Locomotion Robot Locomotion
Tsinghua University · ZERITH Robotics · Shanghai University

Mar 16, 2026J/K move · Enter open · S save

User-Tailored Learning to Forecast Walking Modes for Exosuits

Gabriele Abbate, Enrica Tricomi, Nathalie Gierden, Alessandro Giusti+2

Dalle Molle Institute for Artificial Intelligence (IDSIA), USI-SUPSI, Lugano, Switzerland · Munich Institute for Robotics and Machine Intelligence (MIRMI), Department of Computer Engineering, School of Computation, Information and Technology, Technical University of Munich (TUM), Munich, Germany.

Assistive Robotics Wearable Sensing Online Learning

Open paper AskPDF arXiv

Assistive robotic devices, like soft lower-limb exoskeletons or exosuits, are widely spreading with the promise of helping people in everyday life. To make such systems adaptive to the variety of users wearing them, it is desirable to endow exosuits with advanced perception systems. However, exosuits have little sensory equipment because they need to be light and easy to wear. This paper presents a perception module based on machine learning that aims at estimating 3 walking modes (i.e., ascending or descending stairs and walking on level ground) of users wearing an exosuit. We tackle this perception problem using only inertial data from two sensors. Our approach provides an estimate for both future and past timesteps that supports control and enables a self-labeling procedure for online model adaptation. Indeed, we show that our estimate can label data acquired online and refine the model for new users. A thorough analysis carried out on real-life datasets shows the effectiveness of our user-tailored perception module. Finally, we integrate our system with the exosuit in a closed-loop controller, validating its performance in an online single-subject experiment.

Abstract

Explore similar work

User-Tailored Learning to Forecast Walking Modes for Exosuits

GAIT: Legged Robot Proprioceptive State Estimation with Attention over Inertial-Leg Tokens

PRIOR: Perceptive Learning for Humanoid Locomotion with Reference Gait Priors