Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MOB-Net: Limb-modularized Uncertainty Torque Learning of Humanoids for Sensorless External Torque Estimation

Daegyu Lim, Myeong-Ju Kim, Junhyeok Cha, Jaeheung Park

TL;DR

MOB-Net addresses the challenge of estimating pure external joint torque on floating-base humanoids without additional sensors by combining a model-based momentum observer with limb-modularized data-driven networks that learn the system's uncertainty torque. The six limb-group GRU-Linear networks produce a probabilistic estimate of the uncertainty torque that calibrates the MOB residual to yield accurate external torque estimates in real time. The approach is validated through extensive simulations and real-humanoid experiments, showing substantial improvements over MOB, friction models, and end-to-end learning in both in-distribution and unseen data, and enabling sensorless collision detection and reaction. This work offers a cost-effective path to safer, more robust humanoid operation by enabling full-body collision handling using proprioceptive sensing alone, with potential extensions to control and sim-to-real transfer.

Abstract

Momentum observer (MOB) can estimate external joint torque without requiring additional sensors, such as force/torque or joint torque sensors. However, the estimation performance of MOB deteriorates due to the model uncertainty which encompasses the modeling errors and the joint friction. Moreover, the estimation error is significant when MOB is applied to high-dimensional floating-base humanoids, which prevents the estimated external joint torque from being used for force control or collision detection in the real humanoid robot. In this paper, the pure external joint torque estimation method named MOB-Net, is proposed for humanoids. MOB-Net learns the model uncertainty torque and calibrates the estimated signal of MOB. The external joint torque can be estimated in the generalized coordinate including whole-body and virtual joints of the floating-base robot with only internal sensors (an IMU on the pelvis and encoders in the joints). Our method substantially reduces the estimation errors of MOB, and the robust performance of MOB-Net for the unseen data is validated through extensive simulations, real robot experiments, and ablation studies. Finally, various collision handling scenarios are presented using the estimated external joint torque from MOB-Net: contact wrench feedback control for locomotion, collision detection, and collision reaction for safety.

MOB-Net: Limb-modularized Uncertainty Torque Learning of Humanoids for Sensorless External Torque Estimation

TL;DR

MOB-Net addresses the challenge of estimating pure external joint torque on floating-base humanoids without additional sensors by combining a model-based momentum observer with limb-modularized data-driven networks that learn the system's uncertainty torque. The six limb-group GRU-Linear networks produce a probabilistic estimate of the uncertainty torque that calibrates the MOB residual to yield accurate external torque estimates in real time. The approach is validated through extensive simulations and real-humanoid experiments, showing substantial improvements over MOB, friction models, and end-to-end learning in both in-distribution and unseen data, and enabling sensorless collision detection and reaction. This work offers a cost-effective path to safer, more robust humanoid operation by enabling full-body collision handling using proprioceptive sensing alone, with potential extensions to control and sim-to-real transfer.

Abstract

Momentum observer (MOB) can estimate external joint torque without requiring additional sensors, such as force/torque or joint torque sensors. However, the estimation performance of MOB deteriorates due to the model uncertainty which encompasses the modeling errors and the joint friction. Moreover, the estimation error is significant when MOB is applied to high-dimensional floating-base humanoids, which prevents the estimated external joint torque from being used for force control or collision detection in the real humanoid robot. In this paper, the pure external joint torque estimation method named MOB-Net, is proposed for humanoids. MOB-Net learns the model uncertainty torque and calibrates the estimated signal of MOB. The external joint torque can be estimated in the generalized coordinate including whole-body and virtual joints of the floating-base robot with only internal sensors (an IMU on the pelvis and encoders in the joints). Our method substantially reduces the estimation errors of MOB, and the robust performance of MOB-Net for the unseen data is validated through extensive simulations, real robot experiments, and ablation studies. Finally, various collision handling scenarios are presented using the estimated external joint torque from MOB-Net: contact wrench feedback control for locomotion, collision detection, and collision reaction for safety.
Paper Structure (35 sections, 19 equations, 22 figures, 12 tables)

This paper contains 35 sections, 19 equations, 22 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Model Uncertainty Handling in External Torque Estimation
  4. Collision Handling for Legged Robot
  5. Comparison to our previous work
  6. Problem Formulation
  7. Rigid Body Dynamics of Floating-base Robot
  8. Momentum Observer (MOB)
  9. Uncertainty Torque
  10. Problem Definition
  11. Model Uncertainty Torque Learning for Pure External Torque Estimation
  12. MOB-Net Architecture
  13. Input Feature and Target Value
  14. Quality of Information
  15. Potential Target Task
  16. ...and 20 more sections

Figures (22)

  • Figure 1: An example of collision detection on the foot of a humanoid robot with a heavy box due to the occlusion of the camera sensor on the head and collision reaction by back-stepping. See Scenario #3 in https://youtu.be/ZpnMEjvGsaQ for the video.
  • Figure 2: Diagram of the proposed method. (a) Structure of the inertia matrix of a humanoid robot. The empty region has zero elements. (b) The architecture of the proposed MOB-Net. Each color bar represents the input data from the specific limb.
  • Figure 3: Random walking data collection environment for the simulation and the real robot experiment. (a) Simulation environment with uneven terrain of check pattern. (b) Uneven terrain setting with three wooden plates of approximately 10mm thickness. (c) Slope terrain setting of approximately 3 degrees.
  • Figure 4: A humanoid robot, TOCABI, used in the experiment and its joint names. Only the joint names in the left limbs are displayed due to their symmetry.
  • Figure 5: Simulation results of forward walking on uneven terrain. The estimated external joint torque of the LL3 (left hip pitch) joint is plotted according to the uncertainty level and the estimation method. The robot walks forward on uneven terrain with a step duration of 0.7 s and the plots show one cycle of the locomotion.
  • ...and 17 more figures

Theorems & Definitions (1)

  • Definition 3.1