WEBVTT

1
00:00:00.110 --> 00:00:04.549
We present cross-embodied co-design for dexterous robot hands.

2
00:00:05.629 --> 00:00:10.769
Traditionally, robot bodies are designed first and a method
for control second.

3
00:00:11.529 --> 00:00:16.210
This can lead to suboptimal designs and control pairs
for challenging tasks.

4
00:00:17.329 --> 00:00:22.410
We show developing control and design together can lead
to improved task performance.

5
00:00:23.269 --> 00:00:29.390
Our co-design framework aims to bridge this gap by
generating robot hands and control together.

6
00:00:29.390 --> 00:00:37.850
We provide a framework to generate, evaluate, and control
robot hands with complementary controls, sim-to-real.

7
00:00:38.609 --> 00:00:48.009
First, we create cross-embodied policies which work across robot
hand variations from a randomly generated robot hand dataset.

8
00:00:48.909 --> 00:00:57.109
Next, we generate and evaluate designs using modular parts,
learned design rules, and design-conditioned policies.

9
00:00:57.109 --> 00:01:02.570
Finally, we manufacture our robot hands and deploy them
sim-to-real.

10
00:01:03.130 --> 00:01:07.810
We deploy across multiple robot hands and three tasks
in simulation.

11
00:01:08.909 --> 00:01:16.569
In-hand rotation: the robot hand must rotate randomized objects
around the x-axis.

12
00:01:17.489 --> 00:01:24.730
Object grasping: this tests the ability to create design
and control with quick and stable grasps.

13
00:01:24.730 --> 00:01:34.590
Object flipping: the robot hand has a fixed wrist
and is asked to rotate randomized objects on the z-axis.

14
00:01:35.250 --> 00:01:47.549
To quickly develop multiple robot hands we created a modular
platform where palms, fingertips, degrees of freedom, finger lengths,
and shapes can be easily altered.

15
00:01:48.250 --> 00:01:50.590
The build guide is available on our website.

16
00:01:51.430 --> 00:02:03.310
We show trade-offs for design and control strategies and
the ability of our framework to accurately predict design ranking
sim-to-real by testing four robot hands on a blind
in-hand rotation.

17
00:02:03.969 --> 00:02:16.930
Top is the best found design for in-hand rotation —
a three-finger design which provides the maximum torque while providing
the minimum points required to define a stable circumference
of rotation.

18
00:02:16.930 --> 00:02:23.370
Bottom is our baseline human-like hand which has five
fingers with a human-like layout.

19
00:02:24.150 --> 00:02:33.569
We show two additional designs sim-to-real to demonstrate the
ability of our framework to predict design-control ranking accurately.

20
00:02:34.110 --> 00:02:38.530
Our paper and robot hand build guide are available
on our website.