12 November 2009, UT Austin, CS Department Control of Humanoid Robots Luis Sentis, Ph.D. Personal robotics Guidance of gait
Assessment of Disruptive Technologies by 2025 (Global Trends)
Human on the loop: Personal / Assitive robotics (health) Unmanned surveillance systems (defense / IT) Modeling and guidance of human movement (health) Human-Centered Robotics
Biomechatronics facility Emerging sensors Analysis of Human Gait (a) (b) (c) (d) Human-Centered Robotics Laboratory
Human Centered Robotics Today
Current Projects: Compliant Control of Humanoid Robots
Recent Project: Guidance of Gait Using Functional Electrical Stimulation
CONTROL OF HUMANOID ROBOTS
General Control Challenges Dexterity: How can we create and execute advanced skills that coordinate motion, force, and compliant multi-contact behaviors Interaction: How can we model and respond to the constrained physical interactions associated with human environments? Autonomy: How can we create action primitives that encapsulate advance skills and interface them with high level planners PARKOUR
The Problem (Interactions) Operate efficiently under arbitrary multi-contact constraints Respond compliantly to dynamic changes of the environment Plan multi-contact maneuvers Coordination of complex skills using compliant multi-contact interactions
Key Challenges (Interactions) Find representations of the robot internal contact state Express contact dependencies with respect to frictional properties of contact surfaces Develop controllers that can generate compliant whole-body skills Plan feasible multi-contact behaviors
Approach (8 years of development) 1.Models of multi-contact and CoM interactions 2.Methodology for whole-body compliant control 3.Planners of optimal maneuvers under friction 4.Embedded control architecture
Humanoids as Underactuated Systems in Contact Non-holonomic Constraints (Underactuated DOFs) External forces Model-based approach: Euler-Lagrange Torque commands Whole-body Accelerations External Forces
Model of multi-contact constraints Accelerations are spanned by the contact null-space multiplied by the underactuated model: Assigning stiff model:
Model of Task Kinematics Under Multi-Contact Constraints x q legs Reduced contact-consistent Jacobian x base q arms Differential kinematics Operational point (task to joints)
Modeling of Internal Forces and Moments
Variables representing the contact state
Aid using the virtual linkage model (predict what robot can do) C C C C Grasp / Contact Matrix Center of pressure points Internal tensions Center of Mass Normal moments
Properties Grasp/Contact Matrix 1.Models simultaneously the internal contact state and Center of Mass inter- dependencies 2.Provides a medium to analyze feasible Center of Mass behavior 3.Emerges as an operator to plan dynamic maneuvers in 3d surfaces
Example on human motion analysis (is the runner doing his best?)
More Details of the Grasp / Contact Matrix Balance of forces and moments: Underdetermined relationship between reaction forces and CoM behavior: Optimal solution wrt friction forces
Example on analysis of stability regions (planning locomotion / climbing)
Contact Center of Pressures (CoPs) C Balance of moments on support links
Dependency CoP’s – ZMP
Dependency CoP’s – ZMP (Coplanar Stance Only) Relationship CoP’s - ZMP Dependencies
Approach 1.Models of multi-contact and CoM interactions 2.Methodology for whole-body compliant control 3.Planners of optimal maneuvers under friction 4.Embedded control architecture
Linear Control Stanford robotics / AI lab Torque control: unified force and motion control (compliant control) Control of the task forces (pple virtual work) Control of the task motion Potential Fields
Inverse kinematics vs. torque control duality Pros: Trajectory based Cons: Ignores dynamics Forces don’t appear Pros: Forces appear Compliant because of dynamics Cons: Requires torque control Inverse kinematics: Torque control:
Highly Redundant Systems Under Constraints
Prioritized Whole-Body Torque Control Prioritization (Constraints first): Gradient descent is in the manifold of the constraint
Constrained-consistent gradient descent x task Optimal gradient descent: Constrained kinematics: x un-constrained
Constrained Multi-Objective Torque Control Lightweight optimization Decends optimally in constrained-consistent space Resolves conflicts between competing tasks
Torque control of humanoids under contact
Control of Advanced Skills
Example: Interactive Manipulation
Manifold of closed loops Control of internal forces Unified motion / force / contact control
Compliant Control of Internal Forces Using previous torque control structure, estimation of contact forces, and the virtual linkage model:
Simulation results
Approach 1.Models of multi-contact and CoM interactions 2.Methodology for whole-body compliant control 3.Planners of optimal maneuvers under friction 4.Embedded control architecture
Contact Requisites: Avoid Rotations and Friction Slides C Rotational Contact Constraints: Need to maintain CoP in support area Frictional Contact Constraints: Need to control tensions and CoM behavior to remain in friction cones
Automatic control of CoP’s and internal forces Motion control
CoM control
Example: CoM Oscillations
Lateral Walk: CoM and CoP Trajectoriy Generation
Specifications
Multiple steps: forward trajectories
Multiple steps: lateral trajectories
Results: lateral steps
Dynamic Walk
Approach 1.Models of multi-contact and CoM interactions 2.Methodology for whole-body compliant control 3.Planners of optimal maneuvers under friction 4.Embedded control architecture
Cognitive architecture
Demos Asimo Upper body compliant behaviors Honda’s balance controller Torque to position transformer
Manipulation tests
Summary Grasp Matrix 1.Models of multi-contact and CoM interactions 2.Methodology for whole-body compliant control 3.Planners of optimal maneuvers under friction 4.Embedded control architecture
PRESENTATION’S END