We propose Logic Network Flow (LNF), an optimization-based temporal-logic task and motion planning framework that encodes temporal predicates as network flow constraints, yielding provably tighter convex relaxations and computational speedups of up to several orders of magnitude. LNF is validated on various problems such as vehicle routing, multi-robot coordination, hybrid dynamical systems, and real-time replanning on quadrupedal robots.

We propose a model predictive control framework using Generalized Benders Decomposition (GBD), with applications to contact-rich systems such as a cart-pole with soft-contact walls and humanoid balancing assisted by hand contacts, reaching speeds 2-3 times faster than Gurobi and oftentimes exceeding 1000Hz.

We present a Benders Decomposition approach with time-shifted cuts for bipedal task and motion planning under Signal Temporal Logic (STL) constraints. The method partitions the NP-hard hybrid problem into a master problem for task scheduling and subproblems for kinematic and dynamic feasibility checks, achieving up to 20x speedup on logistics and automation scenarios.

We compare data-driven acceleration techniques for mixed-integer bilinear programs (MIBLPs) under two reformulations: mixed-integer convex programs (MICP) via McCormick envelopes, and mathematical programs with complementarity constraints (MPCC). Using KNN-based warm-starting, we benchmark both on a linear inverted pendulum with soft-contact walls and a single rigid body with mode transitions and contacts, demonstrating real-time motion planning for the SCALER robot transitioning between bipedal and quadrupedal configurations.

We use Reinforcement Learning to enable walking and running gaits for the Booster T1 humanoid, then employ Divergent-Component-of-Motion (DCM) as a mid-level footstep planner that account for the different dynamics and step constraints of each locomotion mode.

We introduce SCALER, a quadrupedal robot capable of climbing bouldering walls, overhangs, and ceilings, as well as trotting on the ground while carrying payloads up to 233% of its weight on flat surfaces and 35% on vertical walls. The first author received the IROS 2022 SICE International Young Authors award. Congratulations!

We introduce SiLVIA, a hexapod robot that demonstrates climbing between two walls with bare foot and planned motion, the first robot to demonstrate such capability.