SLIDE 1 Stabilizing Series-Elas0c Point-Foot Bipeds using Whole-Body Opera0onal Space Control
D.H. Kim, G. Thomas, L. Sen0s The Human Centered Robo0cs Lab The University of Texas Aus0n
1
Workshop on Whole-Body Mul0-Task Mul0-Contact Humanoid Control Humanoids 2015, Seoul
SLIDE 2 Context
- DRC extensively used WBCs
- Locomo0on limita0ons
- Speed and accuracy required
- SEA-based humanoid robots
sought for safety and mul0contact
2
SLIDE 3 Goals
Devising a WBOSC strategy for point-foot bipedal robots Formula0ng force feedback control of internal forces Experimen0ng with mo0on and force behaviors over disjointed terrains using WBOSC Formula0ng planning algorithms for achieving unsupported dynamic balancing
3
SLIDE 4 Terminology: Whole-Body Opera0onal Space Control (WBOSC)
- “Whole-Body Control” (WBC) was an older
terminology we used prior to the crea0on of the IEEE-RAS Technical Commi]ee on Whole-Body
- Control. We decided to use “Whole- Body
Opera0onal Space Control”
- WBOSC formulates dynamically-consistent floa0ng-
based priori0zed task mo0on and internal force torque-controllers.
4
SLIDE 5 Basic Equa0ons of WBOSC
Floa0ng base whole-body dynamics with contact supports Support consistent task differen0al kinema0cs Support consistent task differen0al kinema0cs
5
SLIDE 6 (Cont’d)
Task-level accelera0on control structure Internal wrench control structure Mapping between reac0on and internal wrenches
6
SLIDE 7 Equivalent Op0miza0on
7
SLIDE 8 Significance of Work
- WBOSC is newly formulated for elas0c point-foot
bipedal robots
- WBCs can be easily transferred between robots
- Some WBCs decouple task dynamics
- WBOSC enables feedback control of mo0on tasks
and internal force behaviors
- Point-foot bipeds do not have a suppor0ng polygon
8
SLIDE 9 Technical Procedure
- Devise methods to achieve stable and accurate
whole-body opera0onal space
- Feedback control of internal forces for pulling or
pushing on terrains
- IMU-Mocap sensor fusion for pose es0ma0on
- Locomo0on planning and control to stabilize point-
foot robot on various terrains
9
SLIDE 10 Treatment of Contact Forces on WBCs
- Full Body Controller of SARCOS CMU. First solves for
contact forces using approximate QP then solves full- body constrained inverse dynamics problem.
- The Momentum-Based Controller of IHMC has also 2
stages, but solves for joint accelera0ons using the centroidal momentum matrix.
10
SLIDE 11 Treatment of Contact Forces
- n WBOSC
- WBOSC divides contact forces into mo0on tasks and internal forces
- Relies on conver0ng desired contact forces to internal forces
- Contact solver based on mul0contact matrix, or exis0ng QP solvers
- Using feedback to control of internal forces which overcomes
reduce torque control effort
11
SLIDE 12 Pros/Cons WBOSC
Pros:
- Can accomplish accurate contact force control
regardless of joint torque controllers
- Priori0es allows behaviors to automa0cally repress
conflic0ng tasks
- Dynamic correctness
- Supports the execu0on of overdetermined tasks
- Exposes task’s effec0ve iner0a
Cons:
- Necessitates to convert desired contact forces to
internal forces
- Lack of inequality constraints
12
SLIDE 13 SOTA Point Foot Bipeds
- ATRIAS achieves unsupported dynamic walking
without reliance on WBCs. It has not been shown to regulate internal forces on disjointed terrains. It does not control yaw moments, instead relying on small passive feet.
- MARLO achieves 20 steps of unsupported dynamic
- walking. In comparison, Hume achieves 18 steps of
unsupported balance.
13
SLIDE 14 Some Points about Hardware
- Hume, 1.5m, 20Kg, 6 DoF SEAs
– Difficul0es: high s0c0on, flexible structure
- Low latency Microstrain 3DM-GX3-25-OEM
– Difficul0es: large ini0al bias error and noise density
- Phase space impulse mocap
– Difficul0es: suffers from occlusions and large latencies (5ms)
- WBOSC running at 0.667ms, end to end, EtherCat,
RTAI Linux. DSPs perform torque control at 2KHz.
14
SLIDE 15
End to end controller
SLIDE 16
End to end controller
Task Controller
SLIDE 17
End to end controller
SLIDE 18
Opera0onal Space Trajectory
Single Support Double Support
SLIDE 19
End to end controller
SLIDE 20
High Level Posi0on (PID)
PID control
SLIDE 21
End to end controller
SLIDE 22
Contact Switching Transi0on
Double Support Single Support Reac0on Force
SLIDE 23
Experiment Result
SLIDE 24
End to end controller
Internal Force Control
SLIDE 25
Internal Force Control
Internal force:
mutually canceling forces and moments between pairs or groups of contact points, i.e. tensions, compressions and reac0on moments
SLIDE 26
Internal Force Control
Fully controllable - orthogonal to the robot’s mo0on
SLIDE 27 Internal Force Control
Fully controllable - orthogonal to the robot’s mo0on
`
SLIDE 28
How to Design Internal Force
In our experiment, simple geometry is used Op0miza0on based methods (QP) are also possible op0on
SLIDE 29
Internal Forces Experiment
SLIDE 30
Joint Torque Control
SLIDE 31 Problem Defini0on
Reference (given) Measurement
We are finding:
SLIDE 32
Formula0on of The Problem
1st row of
:
Therefore, we can rewrite the equa0ons:
SLIDE 33 Include More Informa0on
- LED occlusion: Ignore some0mes
- How to mix MoCap with IMU data (fast although
noisy)
- Substan0al delay in the MoCap data
Then simply take the pseudo inverse: But we need to think the followings:
SLIDE 34
How to Mix with IMU data
p Previous es0mated value want to find with MoCap & IMU data
SLIDE 35
Sensor Fusion
&
SLIDE 36
Velocity Reversal Foot Placement Planner
SLIDE 37
Unsupported Balance Experiment
SLIDE 38
Thank you for your a]en0on!
SLIDE 39
Impact Model (op0onal)
when (ordinary landing speed),
