Skip to main content
$$ % Define your custom commands here \newcommand{\bmat}[1]{\begin{bmatrix}#1\end{bmatrix}} \newcommand{\E}{\mathbb{E}} \newcommand{\P}{\mathbb{P}} \newcommand{\S}{\mathbb{S}} \newcommand{\R}{\mathbb{R}} \newcommand{\S}{\mathbb{S}} \newcommand{\norm}[2]{\|{#1}\|_{{}_{#2}}} \newcommand{\pd}[2]{\frac{\partial #1}{\partial #2}} \newcommand{\pdd}[2]{\frac{\partial^2 #1}{\partial #2^2}} \newcommand{\vectornorm}[1]{\left|\left|#1\right|\right|} \newcommand{\abs}[1]{\left|{#1}\right|} \newcommand{\mbf}[1]{\mathbf{#1}} \newcommand{\mc}[1]{\mathcal{#1}} \newcommand{\bm}[1]{\boldsymbol{#1}} \newcommand{\nicefrac}[2]{{}^{#1}\!/_{\!#2}} \newcommand{\argmin}{\operatorname*{arg\,min}} \newcommand{\argmax}{\operatorname*{arg\,max}} \newcommand{\dd}{\operatorname{d}\!} $$

Roadmap

Where the program goes from a certified chain: finish the zonotope migration, make the planner compose certified funnels, then learn certificates, use momentum, generalize across scenes, and move toward hardware and 3D.

Where things stand

The planner finds the held indirect lift from the scene alone in 197 of 200 held-out searches (Results). Beside it, one route of the lift, “track C”, is built as a chain of certified funnels: each edge is a transition certified for every state of its entry set and every disk friction in [0.3, 0.5], and the junctions between edges are checked as containments of certified sets (Certified funnels). The chain is executed end to end in MuJoCo/MJX. Its certificates are moving from interval boxes to zonotopes, which carry the correlations that boxes lose (Background: zonotope propagation).

The steps below are ordered by what each one makes possible next.

1. Finish the certified chain and the migration

  • The last measured link. After the landing impact (certified by a zonotope cover of the impact window), the box comes to rest on the platform. That settling is today a measured outcome. It becomes an endpoint certificate started from the impact cover’s certified exit, with the same construction as the chain’s other rests.
  • Every certificate on zonotopes. The funnel certifier gains the zonotope set type edge by edge: a regression shows each migrated edge holds wherever its interval certificate held, with margins at least as large; then track C is re-certified in chain order, so each edge starts from the previous edge’s certified zonotope exit; then the indirect lift’s certified segment. Interval arithmetic stays as the rigorous subroutine for remainder bounds, implicit contact solves and rounding (Background: intervals and zonotopes together).
  • The paper. The method, the certified chain and the set-based certificates, written for a journal (T-RO or IJRR).

2. The planner composes certified funnels

This is the program’s central step. Today the search returns plans that are held in simulation, and the certified chain is built beside it. The goal is a planner whose output is a certified plan:

  • each edge the search expands is given a certified funnel, certified on demand from the node’s certified entry set and cached by edge and entry set;
  • junctions are composed by containment during the search, so a node is reached only through certified hand-offs;
  • the search’s success then means “certified end to end”, and the realizer’s role shifts from finding held motions to proposing laws for the certifier.

The design question to settle first is certification on demand inside the search against a precomputed library of funnels per edge, or a mix.

3. Built on the certified planner

  • Learned certificate proposals. Once the search certifies edges routinely, it produces data: which laws, hand-offs and set shapes certify. A graph network over the contact-formation graph proposes these ingredients (including zonotope generator templates); the exact verifier accepts or refuses, so soundness never depends on the network. Its value is claimed only through ablations.
  • Motions that need momentum. A second route (“track A”) coasts the box through a wedge of poses where no push can hold it: the box has to be moving. It reuses the reach and impact machinery of track C and is the first plan in which momentum is essential rather than incidental.
  • Families of scenes. Scene parameters, such as the step height, the box’s size and mass, enter the certificates as extra zonotope generators, so one certificate covers a family of scenes rather than a single one.
  • The disk sliding along the box. An edge where the disk slides is parked for a method change; the zonotope tools are the next attempt.

4. Toward hardware and 3D

  • Hybrid zonotopes. At contact switches a set splits by contact mode. Hybrid zonotopes represent such mode-branched sets exactly, which matters more as modes multiply. This line is a natural collaboration with Justin Koeln and Justin Ruths at UT Dallas.
  • Perception. A state estimator replaces the simulator’s state; its error bounds enter the certificates the same way today’s measured disturbances do (at joint 95/95).
  • A real robot. The same loop (plan, certify, execute with feedback) on hardware, with the certificates’ disturbance bounds measured on the robot.
  • 3D. More contact modes and a higher-dimensional state, where set-based certificates are the scalable route and box splitting is not.

The order, and why

1 → 2 → 3 → 4. The learned proposer waits for the certified planner because it needs the planner’s certified edges as data; the momentum route and scene families can proceed in parallel with it. Hardware and 3D build on all of it.

The method at a glance · Certified funnels · Results