How Human-Like Can Atlas Be?

IHMC, MIT, and WPI-CMU


This web page really describes metrics to compare robot and human walking.


Why Make Atlas Human-Like?

(Collins) I don't think it makes sense for Atlas to act like a human in terms of kinematics, cadence, step length, ground reactions, etc. The system is so different from a person that we should expect optimal coordination strategies to be very different as well. I say: walk like a robot.

(Atkeson) You make a good point. If we were working with CHIMP or RoboSimian, I would agree with you.

But we have a biped, and we are so far from optimum anything that we still need to capture the basic features of walking that humanoids will share.

Also, I want to understand human walking...

Here are the current contenders for fast walking in Atlas.

MIT

       0.4 m/s     0.35m steps?   *0.875s/steps??      *calculated
                   (left to right)
IHMC
       0.5 m/s    *0.3m steps      0.6s/step
WPI-CMU
       0.4 m/s    *0.32m steps     0.8s/step
Boston Dynamics (over rough terrain)
       ?           ?               0.5s/step

What do you notice? Most of these gaits are flat footed (IHMC has a hint of a heel strike) and the knees are always very bent. Furthermore, IHMC believes they run out of oil flow capability in the knee joints.

These gaits are all moving the knee way too much. How do we improve this? Be more like a human! Straighten the leg, use the ankle joint much more (heel strike and push off), and tolerate more vertical movement of the pelvis (more like a compass gait).

I don't believe any bipedal robot does anything useful or interesting with its pelvis, which is stupid.

I say: walk like a human. When you get to that point, you can then be more like a robot.

(Pratt) I like the challenge of most human-like. I'm not really interested in optimal for a given robot, especially something that is horribly inefficient like Atlas. ... Nailing down a score function would be great both for this challenge and for the community.

Steve, if you need a good reason why human-like is better than most efficient or other such measure, it is because we need to make them blend in better with humans so that they are harder to detect once Skynet comes online.


Feet Matter

(J. Kim) One big difference between human and current humanoid robots is joint mechanism. We don't have pinned revolute joints in our legs, all of our joints are rolling joints. Depending on the contact surface between bones, rolling joints provide varying rolling (rocking) axis, more stable area for standing, more range of motion and so on. One other big difference is foot mechanism. Human's foot has almost same mechanism with his/her hand but most of humanoids have one rigid plate as its one foot. With this plate, it's not easy to implement heel strike and push off (I prefer to call these as heel rocking and toe rocking). In human walking, heel strike occurs while the other foot's toe is rocking.

If we try this with Atlas, there are two line contacts of feet when heel strike occurs. Atlas might have some support polygon even with line contacts but enough friction is not guaranteed. To have enough friction while double support phase (heel rocking and toe rocking), better mechanism such as bending toe, round deformable heel is necessary. Takanishi lab tried lots of feet mechanisms with WABIAN-2R and the following link is one of them. wabian-2R, More works of WABIAN-2R

Ames lab: AMBER 2

I'm not sure what is typical ZMP walking you mentioned, but ASIMO has heel landing and toe off and its ZMP reference is varying from heel to toe during single support. It is described in IROS 2009 proceedings where you can find 4 papers about ASIMO's walking and running. LINK And its swing foot is not parallel to the ground. LINK I'm not saying that ASIMO's walking looks like human. It still has bent knee and tries to keep the trunk parallel to the ground. But I think the reason is important. From the late 1980s to the early 1990s, Honda tried to make straight knee walking based on human motion data. As shown in the following video, they had quite successful results with their lower limb robots. VIDEO I guess that they gave up straight knee, more dynamic walking, after they had full body robot, because they wanted better control stability and state estimation in human environment. As you know, there are lots of noise in the sensor data while walking, Especially, vision data has more noise/blur because eyes are attached to head which is the furthermost link from the stance foot. They just gave up some similarity to human motion and took advantages by having fewer DoF in the upper body. If you want straight knee and more dynamic pelvic motion, it'll be better to have redundant joints instead of pelvis to spine mechansim, as Petman has more pitch joints. LINK The following paper is an effort for vision-based SLAM in heel-toe walking. A bit related to state estimation. LINK

My suggestion to make Atlas walk more human-like is having better feet. Upgrading all the hardware will not be possible but changing feet is relatively easy.



Human speed

1.3m/s

Human cadence

0.5s/step

Human step length

(left to right) 0.65m
(left to left) 1.3m

Heel strike and toe off

Must have.

Previous comparisons of robots and humans

This paper discusses comparing human and humanoid walking patterns. CGA copy Conference version in IEEEXplore.

Human-like ground reaction forces

Fz for robots is often constant in single support (COM constant height). Human Fz has a characteristic double hump with variations 20-30% above and below body weight (COM accelerates vertically up and down). This double hump also allows larger horizontal foot forces at the beginning and end of single support to remain within the friction cone.

Humans have significant Fx and Fy (horizontal forces and corresponding horizontal COM accelerations). ZMP robots typically don't, and thus have small Fx and Fy. You can match COM excursions rather than horizontal foot forces if that is more convenient to measure or estimate.

Mx and My -> COP: Robots often keep COP in the center of the stance foot in single support. Humans start with it at the heel, and move it to the toe during single support, due to heel strike and toe push off.

Mz I don't have information about foot torque around a vertical axis (yaw torque) for either humans or robots.

Human-like kinematics

Humans almost staighten their leg (knee angle goes to zero) at heel strike.

Human have a lot more hip motion due to longer step. Hip movement is symmetric to cancel yaw moment on torso.

Humans have about a 25degree ankle excursion. H7 in this paper had about a 10degree ankle excursion.

Human feet are NOT kept parallel to the ground. This is what leads to heel strike and toe push off.

There is also pelvic tilt (roll) and yaw rotation in humans. Google "gait determinants" for lots more info. See also this page for some historical revisionism.

Robustness to external perturbations

Robustness to external perturbations (survive X Ns impulse laterally/longitudinally/any direction)

step up/step down

Robustness to model error

Ability to adapt to changed inertial parameters (carrying loads, remove battery dummy). Pick up or drop loads.

Walk on soft ground, elastic vs. plastic