Poor tracking performance given uncertainty.

[jc-bao]

❓ Issue

1. performance in an uncertain environment is much worse than without uncertainty.

plot

🤔 Possible solutions

1. uncertainty curriculum
2. investigate the effectiveness of different parameters.

[jc-bao]

🔧 Trail

🧰 Method: track only task performance

• remove object and only control the drone.
• mainly verify the feasibility of L1 attitude rate controller.

👀 Results

• 😃 Tracking error 10cm Robust RMA
• 😢 Expert performance has no difference with the robust.

🤗 Conclusion

• Lower-level drone control is okay
• Current setting does not utilize expert information.

[jc-bao]

🔧 Trail

🧰 Method: Uncertainty-level curriculum

• start from 0.0/0.3 -> full uncertainty (1.0)

👀 Results

• 😢 curriculum cannot generalize at all. RMA Robust
• 😢 poor performance (50cm tracking error. ) (certain environment performance is 30cm. )

🤗 Conclusion

• The problem mainly lies in environment
  • Environment is too hard?
  • Current modelling is wrong?

[jc-bao]

🔧 Trail

🧰 Method: check rope modelling

• move away from center for a little.

👀 Results

$\zeta=0.75$	$\zeta=0.9$	$\zeta=1.0$	$\zeta=1.1$

🤗 Conclusion

• increase rope zeta to avoid unstable rope oscillation.

[jc-bao]

🔧 Trail

🧰 Method: update rope parameter

• $\zeta$ from 0.75 -> 1.05, $\omega_n$ keep 500rad/s
• disable hook displacement.

code snip

        mass_joint = self.mass_drones * \
            self.mass_obj.unsqueeze(
                1) / (self.mass_drones + self.mass_obj.unsqueeze(1))
        rope_force_drones = mass_joint * \
            ((self.rope_wn ** 2) * rope_disp + 2 *
             self.rope_zeta * self.rope_wn * rope_vel)

👀 Results

• larger damping ratio: tracking error 45 cm. Unstable action still exists. video
• disable hook displacement: unstable rope behavior. video
• 🎥 simulation slow motion 🖼️ plot

🤗 Conclusion

• The sudden tightrope lead to an unstable simulation. (sudden velocity change + bouncing of the drone. )