Transcript The MIT Artificial Intelligence Lab

Legged Robots
Gill Pratt
MIT Artificial Intelligence Laboratory — Research Directions
Why Do Robot Systems Emphasize Stiff
Trajectories Instead of Forces?
• Trajectories are more easily seen than forces
• Most Industrial Robot Tasks (to date) are
Trajectory tasks
– Painting
– Welding
• Humans, even when trying to be stiff, are soft
when walking or doing other tasks
• Our Thesis: Most Natural Robotic Tasks require
“low (mechanical) impedance” thinking
MIT Artificial Intelligence Laboratory — Research Directions
Can Legged Robots Work This Way?
• A: Yes! But requires that forces, as well as
trajectories be considered
• First, this requires actuators that have:
–
–
–
–
Low minimum (mechanical) impedance (i.e. can be soft)
High force fidelity + dynamic range
Robustness to Shock
Energy Storage
MIT Artificial Intelligence Laboratory — Research Directions
But We’re Stuck with Electric-Magnetic
Actuators
• Electric actuators have decent power if run at
high speed, but force/torque is low
– Direct drive is too heavy for autonomous robots.
– Gears are necessary to multiply force/torque and allow
the actuator to run at high speed.
– But gears introduce a number of terrible disadvantages
…
MIT Artificial Intelligence Laboratory — Research Directions
Disadvantages of Gear Reduction
• N2 increase in apparent inertia for N:1 speed
reduction
» Low output impedances are impossible to achieve
• Backlash / Friction
» Output force control has low resolution
» Can be improved with novel gears (e.g. Artisan) but not
inexpensively
• Economical Gear Reductions are intolerant to
shock
– Output teeth break due to single tooth contacts
– Can be improved (e.g. harmonic drive) but not inexpensively
• Poor regeneration (back-drive) efficiency
MIT Artificial Intelligence Laboratory — Research Directions
Our Solution: Series-Elastic Actuators
Bearing
Motor and
Gearbox
Actuator
output
Spring
• Spring in series with motor output
• Spring converts motor position into output force
• Measure spring deflection to control output force
• Series elasticity intentionally used to obtain good force control
MIT Artificial Intelligence Laboratory — Research Directions
Series-Elastic Actuators
(Tendon Elastic)
MIT Artificial Intelligence Laboratory — Research Directions
Series-Elastic Actuators
Linear
Revolute/Stiffening
Torsion spring
Compact DC brushless
MIT Artificial Intelligence Laboratory — Research Directions
Spring Flamingo — Low Impedance
QuickTime™ and a
decompressor
are needed to see this picture.
MIT Artificial Intelligence Laboratory — Research Directions
Walking Algorithms: Motivations for the
Virtual Model Control language
•
•
•
•
(almost) all of us have good physical
intuition
(very) few of us have formal control
intuition
Passive walkers work using physical
components
Q: Can active walker algorithms be
expressed using physical metaphors?
– A: Yes, and they perform
surprisingly well
•
Key Idea: Add Control in Parallel with
natural dynamics of mechanism
MIT Artificial Intelligence Laboratory — Research Directions
Virtual Model Control
Peg-Leg 2-D Walking
• Body Height / Posture maintained via a virtual
wheeled “walker”, regardless of # of legs on
ground
MIT Artificial Intelligence Laboratory — Research Directions
Virtual Model Control
Peg-Leg 2-D Walking
• Speed is Controlled by “food placement” of a
virtual dog-track rabbit instead of “foot
placement”.
• Double-support speed control is possible only
because we have good force control on each leg.
MIT Artificial Intelligence Laboratory — Research Directions
Simulated Pole-Balancing Hexapod
under VMC
QuickTime™ and a
decompressor
are needed to see this picture.
MIT Artificial Intelligence Laboratory — Research Directions
“Spring Turkey” under VMC
QuickTime™ and a
decompressor
are needed to see this picture.
MIT Artificial Intelligence Laboratory — Research Directions
“Spring Flamingo” Walking Over Rough
Terrain Blindly
QuickTime™ and a
decompressor
are needed to see this picture.
MIT Artificial Intelligence Laboratory — Research Directions
Informal Robustness
(see papers for real numbers)
QuickTime™ and a
decompressor
are needed to see this picture.
MIT Artificial Intelligence Laboratory — Research Directions
Recent News:
M2 – A Human Sized Biped Robot
12 Degrees of Freedom
28 kg(62 lbs)
0.97 m(38 in) hip height
MIT Artificial Intelligence Laboratory — Research Directions
M2 Hardware First Steps
QuickTime™ and a
decompressor
are needed to see this picture.
QuickTime™ and a
decompressor
are needed to see this picture.
QuickTime™ and a
decompressor
are needed to see this picture.
MIT Artificial Intelligence Laboratory — Research Directions