It bumps its own shell: measuring where the head can really go

Shortly before one in the morning on 25 October, I filmed six seconds of the robot doing nothing. Its head was tilted about 8 degrees to its left, the antennas uneven, and it no longer reacted to anything. I restarted it.

What happened

The log of my program had nothing to say: it was healthy, it had answered me 25 seconds earlier. The system log of the robot would have told more, but it is not kept across restarts: everything before the restart was gone. Lesson one: I want that log to survive the next time.

What I could see on the video was the head pressed against the body’s shell. The most likely story: a motor pushed against the shell goes into protection and cuts its torque. The head sags and tilts. Only a restart brings it back. That fits everything I saw, but without the lost log it remains a hypothesis.

Who pushes the head into the shell?

I listed every part that moves the head:

1. the maker’s face tracking, which turns the head towards a face, with no limit of its own;
2. my program’s ear, which turned the head towards a sound;
3. the glance before an emotion;
4. the recorded emotions themselves;
5. the start of an emotion: its first position is sent at once, at full speed, from wherever the head happens to be.

The model said one thing

The maker publishes a 3D collision model of the robot. Simulating it, the head should touch the shell at about +55 degrees on its left and −50 to −55 on its right. But the model is made of rough cylinders and boxes, and in it the robot’s sleeping pose sits 9 mm inside the shell. I did not trust it.

The robot said another

So I measured on the robot itself, bare, body fixed, turning the head in 5-degree steps. Detecting a contact took three passes. The first compared the pose asked for with the pose measured, which drifted by 6 degrees even at rest: misleading. The second watched each motor with a threshold of 4 degrees, too low, since a free motor already lags by 2 to 4. The third, the right one, flagged a lag over 6.5 degrees or a jump over 2.5 between two steps.

Head position	Free up to	Contact at
Neutral	−35° on its right, +75° on its left	−40° on its right
Tilted down 15°	−60° and +65°	not reached
Tilted up 15°	not noted	−70° and +70°
Nodding only	±40°	±45°
Rolling only	±50°	not reached
Lowered by 10 mm	±60°	not reached
Raised by 10 mm	not noted	−35° and +30°, the limit of the motors more than the shell

The model was wrong by 10 to 15 degrees: optimistic on the right, pessimistic on the left. And the asymmetry is real: −40 on one side, +75 on the other. Something in the head’s shell or in my assembly is slightly off centre. I still have to look.

Five recorded emotions go beyond −35° on the right. One of them, “curious”, goes to −45°. It played 47 seconds before my video, while face tracking already held the head at −37°.

The guardrails, the same night

At two in the morning, on the robot: the “curious” family removed, the glance limited to 25 degrees instead of 35, the ear to 30 degrees instead of 150. And, in passing, the greeting shortened from 8 seconds to 3.

Next

Removing “curious” is a bandage, and I do not want to lose it: it is one of the robot’s best faces. The real fix is a neck. The head will stay between −30 and +45 degrees of the body, and beyond that the body will turn and take the rest. Then the emotions will start softly instead of jumping, a watchdog will back off when a motor strains, and “curious” will come back.