Dexterous Manipulation

MIT 6.881: Robotic Manipulation

Fall 2020, Lecture 21

Follow live at https://slides.com/russtedrake/fall20-lec21/live

(or later at https://slides.com/russtedrake/fall20-lec21)

Shadow robotics dexterous hand

Adroit Hand (as seen in MuJoCo)

Wonik Allegro hand

LUKE Radial Hand

DLR-HIT Hand II

Robonaut Hand

The "Salisbury Hand".  @ 1982

Ken's short biography

OpenAI - Learning Dexterity

“For the Rubik’s cube task, we use 8 × 8 = 64 NVIDIA V100 GPUs and 8 × 115 = 920 worker machines with 32 CPU cores each.  … The cumulative amount of experience ... is roughly 13 thousand years.”

Solving Rubik's Cube with a Robot Hand by OpenAI, arXiv:1910.07113

Solving Rubik's Cube with a Robot Hand by OpenAI, arXiv:1910.07113

Inverse kinematics as an optimization

subject to:

• rich end-effector constraints
• joint limits
• collision avoidance
• "gaze constraints"
• "feet stay put"
• balance (center of mass)
• ...

Take-aways

• Robot hands will get better; but they are not the bottleneck
  • Cost / reliability is admittedly a limitation
• Simulation will get better; but sim2real is not the bottleneck

• Need more progress on planning/control through contact
  • For changes in parameters that do not make or break new contact, gradients could be well behaved; but it comes down to numerics
  • For making/breaking contact, need more
    • Stochastic optimization, contact implicit, mixed-integer, ...

https://forms.gle/N7P1rf6xuYPVu6wE6​