Research Highlights

Pang

Planning and Control for Contact-rich Robotic Manipulation

  • RRT and Trajectory optimization for dexterous hands (>16DOF).
    • RRT takes less than a minute.
    • Trajectory optimization is fast, and only needs trivial initialization.
    • Our tasks are challenging for existing model-based methods.
      • Contact-implicit traj-opt: huge NLP with many local minima.
      • Hybrid dynamics / Mixed-integer: too many contact modes.
  • Impressive demo.
  • Lots of offline computation.
  • Policy is hard to interpret.

Why did our method work?

  • a Convex, Quasi-dynamic Differentiable Simulator
    • \(F = ma\) --> \(F = 0\) (almost)
      • Considers only transitions between equilibriums, ignores transient effects such as damping.
      • Achieves stable integration with much larger time steps. (0.001s --> 0.4s) 
    • Convex approximation of contact dynamics (Anitescu).
      • Easier than Nonlinear Complementarity Programs (NCPs).
      • The extent of relaxation is controlled by a single hyperparameter.
    • Differentiable by differentiating through the KKT conditions of a convex program.
  • Randomized Smoothing --> Deterministic Smoothing.
    • Contact dynamics is non-smooth (\(C^0\)). 
    • We showed sampling, which RL does a lot, makes contact dynamics smoother (\(C^1\)).
    • Our differentiable simulator can be smoothed without sampling, thereby accelerating algorithms that use smoothed gradients.​

 

  • ​The planned trajectories are already stabilizable in Drake!
    • ​which means our simplification of contact dynamics is reasonable.

Before my PhD defense...

Stabilize this with linear control / MPC!

After my PhD defense... dexterous manipulation of arbitrary rigid objects

Robust planning and control.

  • Robust planning/control
    • Different object sizes / friction coefficients
  • Perception /control based-on tactile sensors
    • No occlusion.
    • Enable force feedback.

Reinforcement Learning.

  • Marco Hutter's work on Anymal has been a great inspiration. 
    • Trying their strategy, but with tweaks for manipulation.
      • Cannot use cyclic policy for leg motions.
      • Need to consider information about object?
  • Can our model-based approach lead to easier-to-learn policy parameterization?
    • RRT does sampling, but only remembers the samples. Can learning offer a better way to use the samples?
    • Instead of learning joint angle commands, can we learn delta object pose, and use a robust model-based controller to achieve the delta pose?

State representation for dexterous manipulation.

  • When we close our eyes with an object in our hand, we don't need the pose of the object to locally manipulate it.
  • Are local contact patches sufficient?

Turning the ball by 30 degrees: open-loop

  • Simulated in Drake with the same SDF as the quasi-dynamic model used for planning.

Turning the ball by 30 degrees: open-loop with disturbance

  • Initial position of the ball is off by 3mm.

Turning the ball by 30 degrees: closed-loop with disturbance

  • Initial position of the ball is off by 3mm.
  • Simple controller:
\begin{aligned} \underset{q^\mathrm{u}_+, u}{\min} &\|q_+^\mathrm{u} - q_+^\mathrm{u, ref}\|_\mathbf{Q} + \|u - u^\mathrm{ref} \|_\mathbf{R}, \text{s.t.}\\ & q^\mathrm{u}_+ = \mathbf{A}_\rho^\mathrm{u} q^\mathrm{u} + \mathbf{B}_\rho^\mathrm{u} u \end{aligned}

Smooth linearization

Task: turning the ball by 30 degrees.