Add touchdown detection to build a better bounding gait.

[Nate711]

No description provided.

[maxsu]

@Nate711 how do you envision touchdown detection?

I have an idea involving a gait state machine (GSM) with a per-leg with two states {contact, free} and two transitions {touchdown, takeoff}:

The takeoff transition is triggered in software. The touchdown transition requires a touchdown event detection (TD) analysis over the motor's kinematic signal:

^{Fig. 2 a) Vh is a height velocity signal for the leg. Event detection happens via sliding window regression, and incurs a TD delay: T_d. b) Motor kinematic signal is transformed into trigger signal for gait state machine.}

Challenges / Questions

1. Control architecture. Is there a better mechanism than a locally triggered state machine?
2. Fault recovery. Environment variability and signal noise can cause TD uncertainty
   • Can we infer and compensate for missed events and unexpected environments?
   • May require a more complex GSM with global states.
3. TD delay. Does the TD cycle provide timely & actionable information during the gait cycle?
4. Compute load. Can the Teensy uC carry the additional TD & control logic?
5. Force feedback. Can we leverage the QDD force feedback signal directly?
   • Does it obviate/simplify the V^h transform?
   • Can it increase TD reliability? Reduce compute load?
6. Simulation. Can we simulate this design in-silico?
   • I don't know doggo's current digital twin strategy.

Dynamical TD

Does doggo currently use touchdown detection to stabilize itself during landing? For example, this could dynamically detect and correct tipping (pitch, roll). This seems to require an altogether more complicated touchdown detection and management strategy (TDM). If this exists / if this is a goal, gait TD could be designed as a subsystem of the Jumping TDM code and runtime.

Conclusion/Next Steps

A signal driven state machine TD strategy may be lightweight enough to run on the teensy. I will look for info on fault recovery and alternative systems. More complex TD can inform a stabilization strategy for jump touchdown; such a system can conceivably provide gait TD.

I don't currently have a doggo, so might need to collaborate with a doggo owner (or wait till I build one). Insights for in-silico/digital twin design validation could enable me to contribute without in-loop physical validation. Feedback and alternative design and collaboration suggestions are welcome.