Home   |
 
Robot Locomotion Group 
 
 
The goal of our research is to build machines which exploit their natural dynamics to achieve extraordinary agility, efficiency, and robustness using rigorous tools from dynamical systems, control theory, and machine learning. Our current focus in on robotic manipulation, because the revolution in recent machine learning has opened a pathway in these applications to merging control theory and perception at a level that has never been considered before; ideas like "intuitive physics" and "common-sense reasoning" will meet with rigorous ideas like "model-order reduction" and "robust/adaptive control". It's going to be a great few years! Our previous projects have included dynamics and control for humanoid robots, dynamic walking over rough terrain, flight control for aggressive maneuvers in unmanned aerial vehicles, feedback control for fluid dynamics and soft robotics, and connections between perception and control. The Robot Locomotion Group is a part of Robotics @ MIT and CSAIL.  
Locomotion Group Paper and Multimedia News
The Shortest-Path Problem in Graph of Convex Sets (SPP in GCS) is a recently developed optimization framework that blends discrete and continuous decision making. Many relevant problems in robotics, such as collision-free motion planning, can be cast and solved as an SPP in GCS, yielding lower-cost solutions and faster runtimes than state-of-the-art algorithms. In this paper, we are motivated by motion planning of robot arms that must operate swiftly in static environments. We consider a multi-query extension of the SPP in GCS, where the goal is to efficiently precompute optimal paths between given sets of initial and target conditions. Our solution consists of two stages. Offline, we use semidefinite programming to compute a coarse lower bound on the problem's cost-to-go function. Then, online, this lower bound is used to incrementally generate feasible paths by solving short-horizon convex programs. For a robot arm with seven joints, our method designs higher quality trajectories up to two orders of magnitude faster than existing motion planners. To Appear at WAFR 2024 Faster Algorithms for Growing Collision-Free Convex Polytopes in Robot Configuration Space
We propose two novel algorithms for constructing convex collision-free polytopes in robot configuration space. Finding these polytopes enables the application of stronger motion-planning frameworks such as trajectory optimization with Graphs of Convex Sets and is currently a major roadblock in the adoption of these approaches. In this paper, we build upon IRIS-NP (Iterative Regional Inflation by Semidefinite and Nonlinear Programming) to significantly improve tunability, runtimes, and scaling to complex environments. IRIS-NP uses nonlinear programming paired with uniform random initialization to find configurations on the boundary of the free configuration space. Our key insight is that finding near-by configuration-space obstacles using sampling is inexpensive and greatly accelerates region generation. We propose two algorithms using such samples to either employ nonlinear programming more efficiently (IRIS-NP2 ) or circumvent it altogether using a massively-parallel zero-order optimization strategy (IRIS-ZO). We also propose a termination condition that controls the probability of exceeding a user-specified permissible fraction-in-collision, eliminating a significant source of tuning difficulty in IRIS-NP. We compare performance across eight robot environments, showing that IRIS-ZO achieves an order-of-magnitude speed advantage over IRIS-NP. IRIS-NP2, also significantly faster than IRIS-NP, builds larger polytopes using fewer hyperplanes, enabling faster downstream computation. To Appear at ISRR 2024 GCS*: Forward Heuristic Search on Implicit Graphs of Convex Sets
We consider large-scale, implicit-search-based solutions to Shortest Path Problems on Graphs of Convex Sets (GCS). We propose GCS*, a forward heuristic search algorithm that generalizes A* search to the GCS setting, where a continuous-valued decision is made at each graph vertex, and constraints across graph edges couple these decisions, influencing costs and feasibility. Such mixed discrete-continuous planning is needed in many domains, including motion planning around obstacles and planning through contact. This setting provides a unique challenge for best-first search algorithms: the cost and feasibility of a path depend on continuous-valued points chosen along the entire path. We show that by pruning paths that are cost-dominated over their entire terminal vertex, GCS* can search efficiently while still guaranteeing cost-optimality and completeness. To find satisficing solutions quickly, we also present a complete but suboptimal variation, pruning instead reachability-dominated paths. We implement these checks using polyhedral-containment or sampling-based methods. The former implementation is complete and cost-optimal, while the latter is probabilistically complete and asymptotically cost-optimal and performs effectively even with minimal samples in practice. We demonstrate GCS* on planar pushing tasks where the combinatorial explosion of contact modes renders prior methods intractable and show it performs favorably compared to the state-of-the-art. Project website: https://shaoyuan.cc/research/gcs-star/ Supplemental materials: https://shaoyuan.cc/research/gcs-star/ To Appear at WAFR 2024 Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion
This paper presents Diffusion Forcing, a new training paradigm where a diffusion model is trained to denoise a set of tokens with independent per-token noise levels. We apply Diffusion Forcing to sequence generative modeling by training a causal next-token prediction model to generate one or several future tokens without fully diffusing past ones. Our approach is shown to combine the strengths of next-token prediction models, such as variable-length generation, with the strengths of full-sequence diffusion models, such as the ability to guide sampling to desirable trajectories. Our method offers a range of additional capabilities, such as (1) rolling-out sequences of continuous tokens, such as video, with lengths past the training horizon, where baselines diverge and (2) new sampling and guiding schemes that uniquely profit from Diffusion Forcing's variable-horizon and causal architecture, and which lead to marked performance gains in decision-making and planning tasks. In addition to its empirical success, our method is proven to optimize a variational lower bound on the likelihoods of all subsequences of tokens drawn from the true joint distribution. Project website: https://boyuan.space/diffusion-forcing/ Supplemental materials: https://twitter.com/BoyuanChen0/status/1808538170067407264 To Appear at NeurIPS 2024 Graphs of Convex Sets with Applications to Optimal Control and Motion Planning
This thesis introduces a new class of problems at the interface of combinatorial and convex optimization. We consider graphs where each vertex is paired with a convex program, and each edge couples two programs through additional convex costs and constraints. We call such a graph a Graph of Convex Sets (GCS). Over a GCS we can formulate any optimization problem that we can formulate over an ordinary weighted graph, with scalar costs on the vertices and edges. In fact, for any fixed choice of the variables in the convex programs, a GCS reduces to a weighted graph where we can seek, e.g., a path, a matching, a tour, or a spanning tree of minimum cost. The challenge in a GCS problem lies in solving the discrete and the continuous components of the problem jointly. By combining the modelling power of graphs and convex optimization, GCSs are a flexible framework to formulate and solve many real-world problems. The graph and the combinatorial goal (e.g., finding a path or a tour) model the high-level discrete skeleton of a problem. The convex costs and constraints fill in the low-level continuous details. The primary contribution of this thesis is an efficient and unified method for solving any GCS problem. Starting from an integer-linear-programming formulation of an optimization problem over a weighted graph, this method formulates the corresponding GCS problem as an efficient Mixed-Integer Convex Program (MICP). This MICP can then be solved to global optimality using common branch-and-bound solvers, or approximately by rounding the solution of its convex relaxation. Importantly, both the formulation of the MICP and its solution are fully automatic, and a user of our framework does not need any expertise in mixed-integer optimization. We first describe the GCS framework and the formulation of our MICP in general terms, without presupposing the specific combinatorial problem to be solved over the GCS. We illustrate our techniques through multiple examples spanning logistics, transportation, scheduling, navigation, and computational geometry. Then we focus on the Shortest-Path Problem (SPP) in GCS. This problem is particularly interesting since it generalizes a wide variety of multi-stage decision-making problems and, using our techniques, it can be solved very effectively. We consider two main applications of the SPP in GCS: optimal control of dynamical systems and collision-free motion planning. In these two areas, our techniques either generalize or significantly improve upon algorithms and optimization methods that have been developed for decades and are widely used in academia and industry. Lastly, the techniques introduced in this thesis are implemented in the software packages Drake and gcspy. The former is a large and mature software for robotics. It is open-source and widely used by the community. The second is a very simple and lightweight Python package which is also open source. In this thesis, we will illustrate the usage of gcspy through multiple basic examples.
 
Locomotion Group News
April 22, 2024. PhD Defense. Congratulations to Terry Suh for successfully defending his PhD thesis! April 22, 2024. PhD Defense. Congratulations to Tobia Marcucci for successfully defending his PhD thesis! September 19, 2023. Video. Check out our work at TRI on Diffusion Policy (towards Large Behavior Models) and the corresponding blog post
June 14, 2023. Award. Congratulations to Terry Suh, Max Simchowitz, and Kaiqing Zhang who's paper titled "Do differentiable simulators give better policy gradients?" was recognized with the "Best Paper (of the year) Award" from the IEEE RAS Technical Committee on Model-based Optimization for Robotics. January 20, 2023. PhD Defense. Congratulations to Tao Pang for successfully defending his PhD thesis! July 15, 2022. Award. Congratulations to Terry Suh, Max Simchowitz, and Kaiqing Zhang who's paper titled "Do differentiable simulators give better policy gradients?" was recognized with the "Outstanding Paper Award" at ICML 2022. January 13, 2022. Award. Congratulations to Alexandre Amice, Hongkai Dai, Pete Werner, and Annan Zhang who's paper titled "Finding and Optimizing Certified, Collision-Free Regions in Configuration Space for Robot Manipulators" was recognized with the "Outstanding Paper Award" at WAFR 2022. June 17, 2022. PhD Defense. Congratulations to Yunzhu Li for successfully defending his PhD thesis! June 17, 2022. PhD Defense. Congratulations to Greg Izatt for successfully defending his PhD thesis! August 15, 2020. Talks on Zoom. For better or worse, most research talks these days are now online. I've posted a handful of links to new talks, including Russ on Lex Fridman's AI Podcast, and at the IFRR Colloquium on the Roles of Physics-Based Models and Data-Driven Learning in Robotics. July 20, 2020. PhD Defense. Congratulations to Lucas Manuelli for successfully defending his PhD thesis! May 29, 2020. PhD Defense. Congratulations to Shen Shen for successfully defending her thesis! September 18, 2019. PhD Defense. Congratulations to Twan Koolen for successfully defending his thesis! August 19, 2019. PhD Defense. Congratulations to Pete Florence for successfully defending his thesis! June 27, 2019. Video. Checkout some of the work being done by TRI's manipulation team
October 15, 2018. PhD Defense. Congratulations to Robin Deits for successfully defending his thesis! October 3, 2018. Award. Congratulations to Pete Florence and Lucas Manuelli whose paper Dense Object Nets: Learning Dense Visual Object Descriptors By and For Robotic Manipulation won the Conference Best Paper Award at CoRL 2018! September 19, 2018. Award. Congratulations to Pete Florence and Lucas Manuelli whose paper Dense Object Nets: Learning Dense Visual Object Descriptors By and For Robotic Manipulation won the first ever Amazon Robotics Best Technical Paper Award (2018). June 18, 2018. Award. Congratulations to Ani Majumdar whose paper Funnel libraries for real-time robust feedback motion planning won the first ever International Journal of Robotics Research Paper of the Year (2017).
|