Proprioceptive Torque Sensing in Robot Legs — PatSnap Eureka
Proprioceptive Torque Sensing in Robot Legs: Terrain-Adaptive Walking Explained
Joint torque and force sensing embedded in robot legs — proprioception — lets legged robots detect contacts, estimate terrain geometry, and adapt gait in real time without cameras or LIDAR. Explore 50+ patents from Honda, Boston Dynamics, Ghost Robotics, Tencent, and more.
Five Technical Clusters in Proprioceptive Terrain Adaptation
The patent landscape divides into five overlapping clusters spanning foundational ZMP control through learning-augmented neural policies.
Why Proprioceptive Torque Sensing Is the Core of Terrain Adaptation
Proprioceptive torque sensing — the measurement of forces and torques at robot leg joints — provides the foundational signal for terrain-adaptive walking. Unlike exteroceptive approaches that depend on cameras or LIDAR, proprioceptive data is always available, cannot be blinded by dust or darkness, and operates at the millisecond timescales required for reactive gait control.
Honda's compliance control paradigm, established in a 1994 Japanese patent, set the template still in use today: an estimated tilt of the virtual plane containing the tread location is computed through kinematics from attitude tilt detection values and joint displacement detection values, and the compliance model output is compared to actual measured floor reaction force moments to correct the desired gait. The foot behaves as a spring-damper against unknown terrain, absorbing geometric errors through controlled deformation.
The patent landscape — spanning more than 50 disclosures across US, EP, JP, WO, and CN jurisdictions from the mid-1990s to 2026 — reveals five overlapping technical clusters. Honda accounts for roughly a third of all cited references. The most recent wave of active filings (2020–2026) comes from Chinese research institutions and technology companies including Tencent, Xiaomi, Shanghai Jiao Tong University, Huazhong University, Wuhan University, and Fudan University, with a strong emphasis on quadruped robots and reinforcement learning.
Understanding this landscape is essential for R&D teams working on competitive intelligence in legged robotics, where the pace of patent activity has accelerated sharply since 2020.
Key Metrics from the Proprioceptive Robotics Patent Landscape
All data derived from analysis of 50+ patents via PatSnap Eureka. Filing dates range from 1994 to 2026 across five jurisdictions.
Patent Volume by Key Assignee
Honda dominates with ~15 patents; Tencent leads recent filings with 5+ pending US/EP patents from 2024–2025.
Filing Activity Timeline by Era
From Honda's foundational 1994 work through the Chinese institution surge of 2020–2026, showing the shift toward learning-augmented proprioception.
How Proprioceptive Torque Sensing Enables Terrain Adaptation
Four distinct mechanisms — each grounded in specific patent disclosures — translate raw joint torque data into terrain-adaptive behaviour.
Floor Reaction Force & ZMP-Based Compliance Control
Honda's foundational architecture treats the robot's foot as a spring-damper against unknown terrain. An estimated tilt of the virtual plane at the tread location is computed from kinematics; the compliance model output is compared to actual measured floor reaction force moments to correct the desired gait. The 2013 gait generating device further sets a permissible range for the vertical component of the floor reaction force moment, correcting instantaneous desired motions in real time to maintain yaw stability on sloped or uneven terrain — purely through joint-transmitted force signals, with no camera or LIDAR required.
Spring-damper terrain absorptionSwing-Phase Contact Classification & Reflexive Re-Swing
Joint dynamics of the swing leg are monitored for unexpected torques and compared against odometry estimates to classify whether an impact is a ground touchdown or an obstacle collision. Boston Dynamics' 2022 patent distinguishes knee-strikes against terrain behind the robot body from ground contacts by analysing the pattern of hip, knee, and ankle torques. Ghost Robotics' 2024 leg-stub re-swing algorithm then enacts a two-phase retract-and-re-swing reflex using only joint torque and position data — explicitly designed for environments where vision sensors may be damaged by protruding obstacles.
Vision-independent obstacle recoveryVariable Stiffness Switching & Terrain Slope Estimation
Toyota's approach softens the ankle joint upon touchdown, allowing actual joint angle to deviate from the pre-planned target and conform to the actual surface. The angular difference between actual and target angles serves as a proxy for terrain relief, and the target trajectory is corrected before re-stiffening for support. Huazhong University's quadruped patents take this further: four foot touchdown positions derived from joint encoders through forward kinematics are assembled into over-determined linear equations and solved via quadratic programming least squares to yield the terrain plane slope in real time — no foot-end contact sensors required. Wuhan University's 2021 method additionally detects compliance heterogeneity across the support polygon when front and rear legs contact surfaces of different material stiffness.
Real-time slope estimation from 4 contact pointsFriction Cone-Aware Force Direction & Slip Management
On slippery surfaces, the primary failure mode is lateral foot slide. Google's 2016 slip avoidance patent determines a representation of the coefficient of friction between foot and ground, independently determines the terrain gradient, computes a threshold orientation for the ground reaction force satisfying the friction cone constraint, and adjusts the commanded force direction accordingly — a purely proprioceptive anti-slip strategy. Sony's 2002 disclosure notes that complete elimination of sliding is not achievable or desirable at high speeds; the productive approach is to quantify and manage sliding rather than forbid it. Wuhan University's 2025 biped gait controller combines reinforcement learning with Model Predictive Control, including a lateral foot sliding velocity reward term derived from joint-level data, to predict potential tipping and enter a reduced-frequency slow gait mode before instability occurs.
Friction cone constraint satisfactionFrom Raw Joint Torque to Adaptive Gait: The Proprioceptive Control Flow
A three-stage pipeline transforms sensed torques into terrain-adaptive commands, as described across the patent dataset.
Learning-Augmented Proprioception: Deep Neural Networks & Reinforcement Learning
The most recent generation of terrain-adaptive controllers uses proprioceptive data as the primary input stream to neural network-based policies, often designed to operate when exteroceptive sensors are unavailable.
Tencent's Residual Prediction Architecture (2025)
Tencent Technology's US patent describes a system in which proprioceptive information characterising the motion state of the robot is fed — together with external perception information — into a deep neural network. The network outputs a predicted residual that corrects trajectory generation parameters. Crucially, the architecture separates proprioceptive from exteroceptive streams, allowing the controller to degrade gracefully if the external perception channel is unavailable, relying on the proprioceptive residual prediction alone. This is directly relevant to life sciences and defence robotics applications where sensor reliability cannot be guaranteed.
Environment-Warped Trajectory Optimisation (Tencent America, 2024)
Tencent America's approach warps the ambient space of complex terrain geometry to a flat space using a learned mapping function, then performs contact-aware optimisation in the flat space. Force and rotational variables are pushed forward from the warped space back to the ambient space, enabling physically valid gait trajectories on arbitrary terrain shapes without per-terrain model re-identification. This represents a significant advance in the patent analytics landscape for generalised legged locomotion.
Key Players and Innovation Trends in Proprioceptive Robotics
Honda Motor Co., Ltd. is the single most prolific assignee in this dataset, with at least fifteen distinct patents covering floor reaction force control, gait generation with permissible range constraints, compliance mechanisms, and landing velocity control. Their paradigm consistently treats the six-axis force/torque sensor at the foot as the ground truth signal and builds compliance and ZMP controllers around it. Honda's work spans from the foundational 1994 JP patent through numerous EP and US filings into the 2010s.
Tencent Technology (Shenzhen) / Tencent America is the most active recent filer, with at least five pending US and EP patents from 2024–2025 covering deep neural network-based residual prediction, joint torque control during landing, and environment-warped trajectory optimisation. Their portfolio signals a strong shift toward learning-augmented proprioception, consistent with trends tracked by WIPO in AI-integrated robotics filings.
Boston Dynamics contributes two highly specific patents on proprioceptive impact classification, directly addressing the challenge of distinguishing ground touchdowns from obstacle collisions using only joint dynamics and odometry. Ghost Robotics Corporation contributes the leg-stub re-swing reflex algorithm, explicitly designed for use when vision sensors fail. Google Inc. contributes friction-aware ground reaction force direction control, linking joint torque measurements to surface coefficient-of-friction estimates and cone constraints.
Chinese institutions — including Huazhong University of Science and Technology, Wuhan University, Fudan University, Shanghai Jiao Tong University, and several companies — represent the most recent wave of active filings (2020–2026). The innovation intelligence community has noted this rapid acceleration in Chinese quadruped robotics IP. For teams building data pipelines to monitor this space, automated patent alerts are now essential.
Seven Insights from the Proprioceptive Torque Sensing Patent Landscape
Drawn directly from the 50+ patent dataset, these findings define the current state and trajectory of the field.
Proprioceptive torque sensing is the foundational terrain-adaptive modality
Floor reaction force and joint torque data allow legged robots to detect surface slope, compliance, and contact state without external sensors, as established in Honda's 1994 walk control patent and still central to modern systems. The PatSnap platform traces this lineage across 30+ years of IP.
Contact classification from unexpected joint torques enables obstacle-robust swing phase
Boston Dynamics' 2022 patent demonstrates that patterns of hip, knee, and ankle torques during swing can distinguish ground touchdown from knee-strikes against terrain above or below the robot body, triggering appropriate reflexes in each case. Verified by IEEE robotics literature on reactive gait control.
Proprioceptive reflex re-swing provides terrain traversal without functional vision
Ghost Robotics' 2024 leg-stub re-swing algorithm shows that a two-phase retract-and-re-swing reflex driven purely by torque sensing can recover from unexpected mid-swing contacts in environments where cameras cannot be trusted.
Terrain slope estimated from four touchdown positions via least-squares optimisation
Huazhong University's 2023 quadruped patent shows that four foot contact positions derived from joint encoders through forward kinematics are sufficient to compute the current terrain plane slope in real time via quadratic programming least squares — no dedicated terrain sensors needed.
Variable joint stiffness switching upon touchdown enables passive terrain conformation
Toyota's 2011 patent demonstrates softening ankle joints during landing, measuring the angular deviation from the pre-planned trajectory as a proxy for terrain relief, and subsequently correcting the target trajectory before re-stiffening for support phase.
Learning-based controllers consuming proprioceptive streams represent the current frontier
Tencent's 2025 legged robot control patent and Shenzhen Zhuji Dynamics' 2026 imitation learning method demonstrate that deep neural networks and reinforcement learning with contact-sequence rewards can systematically encode proprioceptive terrain adaptation, reducing sim-to-real transfer gaps and slip instability. Track this frontier via PatSnap Analytics.
Proprioceptive Torque Sensing in Robot Legs — key questions answered
Proprioceptive torque sensing refers to torque and force sensing embedded within robot leg joints that enables legged robots to detect unexpected contacts, classify terrain types, modulate joint stiffness, and maintain dynamic balance without relying solely on exteroceptive sensors such as cameras or LIDAR.
Honda's compliance control architecture makes the robot's foot behave as a spring-damper against an unknown terrain shape, absorbing errors between the assumed and actual floor geometry through controlled deformation. An estimated tilt of the virtual plane containing the tread location is computed through kinematics, and the compliance model output is compared to actual measured floor reaction force moments to correct the desired gait.
Joint dynamics of the swing leg are monitored for unexpected torques and compared against odometry estimates of robot pose to determine whether the impact corresponds to a genuine ground touchdown or a collision with an obstacle above ground level. Classification logic distinguishes between the knee contacting terrain below the robot body versus behind it, enabling different reflexive responses in each case.
Ghost Robotics Corporation's leg-stub re-swing is a two-phase reflex algorithm: when unexpected contact is detected during swing, the algorithm first retracts the leg stub and then re-initiates the swing trajectory toward the intended touchdown location, using only joint torque and position data to determine the appropriate corrective motion. It is explicitly designed for unstructured terrain environments where vision sensors may be damaged upon contact with protruding obstacles.
Huazhong University of Science and Technology's approach assembles foot touchdown positions derived from joint encoders through forward kinematics into a system of over-determined linear equations for terrain plane estimation. Support leg contact points in the current gait cycle and swing leg contact points from the previous cycle together form four measurements that are solved via quadratic programming least squares to yield the slope of the terrain plane in real time.
Google's slip avoidance method determines a representation of the coefficient of friction between the foot and the ground surface, and independently determines the gradient of the ground surface. Based on both quantities, a threshold orientation is computed for the ground reaction force such that the friction cone constraint is satisfied, and the target ground reaction force direction is adjusted to lie within this cone before being commanded to the leg actuators.
Still have questions? Let PatSnap Eureka answer them for you.
Ask PatSnap Eureka Your QuestionAccelerate Your Legged Robotics R&D with AI-Powered Patent Intelligence
Join 18,000+ innovators already using PatSnap Eureka to accelerate their R&D. Explore 50+ proprioceptive robotics patents, track key assignees, and identify white spaces — all in one platform.
References
- Walk Control Device for Leg Type Moving Robot — Honda Giken Kogyo Kabushiki Kaisha, 1994
- Gait Controller for Legged Mobile Robot — Honda Giken Kogyo Kabushiki Kaisha, 2000
- Legged Mobile Robot and Control Program for the Robot — Honda Motor Co., Ltd., 2011
- Legged Mobile Robot and Control Program for the Robot — Honda Motor Co., Ltd., 2007
- Gait Generating Device of Legged Mobile Robot and Control Device of Legged Mobile Robot — Honda Motor Co., Ltd., 2013
- Footprint Contact Detection — Boston Dynamics, Inc., 2022
- Footstep Contact Detection — Boston Dynamics, 2022
- Leg-Stub Re-Swing for Legged Robot — Ghost Robotics Corporation, 2024
- Leg-STUB Re-Swing for Legged Robot — Ghost Robotics Corporation, 2024 (WO)
- Control Method of Legged Mobile Robot and Legged Mobile Robot — Toyota Motor Corporation, 2011
- Legged Mobile Robot and Control Method — Toyota Motor Corporation, 2010
- Quadruped Robot and Slope Terrain Adaptive Motion Method and Control System — Huazhong University of Science and Technology, 2023
- Quadruped Robot and Slope Terrain Adaptive Motion Method and Control System — Huazhong University of Science and Technology, 2024
- Active Leg Adjustment Method for Quadruped Robot in Variable-Stiffness Terrain Stable Transition — Wuhan University of Science and Technology, 2021
- Slip Avoidance — Google Inc., 2016
- Slip Avoidance — Google Inc., 2017
- Legged Robot Control Method, Legged Robot, and Storage Medium — Tencent Technology (Shenzhen) Company Limited, 2025
- Environment Warped Gait Trajectory Optimization for Complex Terrains — Tencent America LLC, 2024
- Robot Motion Planning Method, System, Computer Device and Storage Medium — Guangdong BSST Robotics Co., Ltd., 2020
- Robot Motion Control Strategy Network Training Method Based on Imitation Learning — Shenzhen Zhuji Dynamics Technology Co., Ltd., 2026
- Tactile-Based Robot Foot-Ground Compliance Control Method and Control System — Fudan University, 2024
- Tactile-Based Robot Foot-Ground Compliance Control Method and Control System — Fudan University, 2025
- Robot Gait Control Method Adapted to Slippery Surfaces, Device, and Biped Robot — Wuhan University, 2025
- Quadruped Robot Motion Planning Method for Complex Terrain — Shanghai Jiao Tong University, 2013
- Legged Mobile Robot and Relative Motion Measurement Sensor — Sony Corporation, 2002
- IEEE — Institute of Electrical and Electronics Engineers (robotics literature reference)
- WIPO — World Intellectual Property Organization (AI-integrated robotics patent trends)
All data and statistics on this page are sourced from the references above and from PatSnap's proprietary innovation intelligence platform. Patent analysis conducted via PatSnap Eureka across US, EP, JP, WO, and CN jurisdictions with filing dates from 1994 to 2026.
PatSnap Eureka searches patents and research to answer instantly.