Sim to real - Leg design issue

[wookm5]

@danic85 So, as you know I am currently working on sim to real for the buddy robot but I am facing some issues. I think the legs may have to undergo another redesign as I am finding it difficult to run the policy when load is applied. The current design when under load will move the upper leg instead of the lower leg, I think this is an issue. I took a look at the open duck project since they are using the same servos and they have a leg design that aligns the servo horn directly on the knee joint position instead of having the two gear design approach at the knee. I could be wrong though and maybe I have issues with my policy or the sim representation of my bot. I did leave out the internal components in sim, so maybe that is causing the policy to not learn under the correct load but it seems like a redesign to align the servo horn directly to the knee might be good enough to get the current policy working. Do you have any thoughts?

[wookm5]

I may retrain as well and make sure to add all components to the bot in sim and be more precise with the servo configs but I have seen some walking movement from the bot in real world which makes me want to do a quick redesign and retry the policy.

[danic85]

Hi @wookm5 ! I agree that it adds unnecessary complexity. Also, there is a lot of 'play' in the gears so the sim vs real world is unlikely to match well.
Would it help if I gave you access to the new version legs (Cody)?

They use ST3215 servos and have abandoned the gear configuration so they will eliminate that issue. They also allow for lateral movement.

The feet, body, neck and head are likely to undergo changes, but the legs could be used for training in their current state (assuming you can add whatever configuration is missing to allow an export to urdf or similar).

I'd need to add your account to the Onshape design as it's currently private, so send me a message on this form and I'll add you.

I'm also happy to have a session with you to discuss if that would be helpful.

[wookm5]

@danic85 Hmmmm, I like the new leg design, and having access to it could definitely help. That said, to avoid any potential negative impact when transferring the policy or the need to retrain I’d prefer to stay as close to the current design as possible. I do have an idea that I think could work, and I’d be happy to discuss it with you whenever you’re available. Also, I’m not sure if this was the correct place to reach out, but I did send a message through the Partners – Get In Touch link as well.

[danic85]

No problem, I've sent you an email. 👍

[wookm5]

So, I did switch the design to position servos horns at joints and I still see the same behavior so its safe to say that the joint position alone is not the reason for my issue. In my novice opinion I think I will have to add another DOF at the hips and retrain.

[danic85]

Interesting! Are you thinking of the hip transverse plane (to rotate the leg left and right)?
I was aware that is missing but I was hoping it would only inhibit turning.

[wookm5]

Well, I don’t think the lack of additional hip DOF alone should prevent walking, but with these servos maybe it does. They simply don’t seem to have enough torque to reliably lift the leg. I tested a single leg in isolation, and even when it was detached from the body and placed on the ground, the motion still propagated up the chain, ankle actuation moved the lower leg, the knee shifted the upper leg, and the hip motion transferred directly into the body.

My thinking was that adding hip motion could help distribute or relieve some of the load on each leg during walking, and it looks like the Duck project’s gait is doing something similar. At this point, it may come down to a combination of torque limitations and training.

I’m planning to try running the existing policy with the new leg design first. If that doesn’t work, I’ll retrain using more precise and realistic servo configurations. To reiterate, the robot should still be able to walk without hip DOF. I’ve seen similar biped builds with comparable gait patterns walk successfully but those designs typically use much more powerful actuators.