Russ Tedrake: Textbooks, Drake and Large Behavior Models

Roles as stated on official pages

Russ Tedrake's MIT homepage gives his title as Toyota Professor of Electrical Engineering and Computer Science, Aeronautics and Astronautics, and Mechanical Engineering. The AeroAstro faculty listing adds that he directs the Robot Locomotion Group at CSAIL. The group page describes its goal as building machines that exploit their natural dynamics, with a current focus on robotic manipulation that combines control theory and perception. [1] [2] [3]

His homepage says he spent ten years as Senior Vice President of Robotics Research and Large Behavior Models at the Toyota Research Institute, in the past tense, and that he founded a Physical AI startup. A Business Wire release of 15 July 2026 names that company: Walden Robotics, with Tedrake as co-founder and CEO, launched out of TRI in January 2026 with a 300 million dollar seed round that the release says was co-led by Toyota entities and Deviation Capital. The CoRL 2026 keynote page uses the same description. The funding figure is the company's statement. [1] [4] [5]

TRI's own people page had not been updated on 19 October 2026 and still shows him as SVP, Large Behavior Models. A reader citing his affiliation should use the MIT professorship, which every page agrees on, and treat the TRI title as former. [6] [1]

Two earlier facts are documented by MIT itself. He led Team MIT's entry in the DARPA Robotics Challenge, which placed sixth of 25 teams in June 2015 with a Boston Dynamics Atlas. And in August 2025 TRI named him co-lead, with Scott Kuindersma of Boston Dynamics, of the Atlas Large Behavior Model project that began in October 2024. [7] [8]

Underactuated Robotics: 21 chapters built around Drake

Underactuated Robotics: Algorithms for Walking, Running, Swimming, Flying, and Manipulation is the online textbook of MIT course 6.8210. The site shows a 2024 copyright and leaves its last-modified field blank, so a revision after 2024 cannot be confirmed from the page. The HTML version is described as the main version; a PDF download is offered. [9] [10]

The book runs from fully actuated versus underactuated systems, through the simple pendulum, acrobots, cart-poles and quadrotors, simple models of walking and running, and legged robots with many joints, into dynamic programming, linear quadratic regulators, Lyapunov analysis, trajectory optimization, policy search, sampling-based motion planning, robust and stochastic control, feedback motion planning, output feedback, limit cycles, planning and control through contact, system identification, state estimation, model-free policy search and imitation learning. Five appendices cover Drake, multibody dynamics, optimization and mathematical programming, an optimization playbook and miscellany. [9]

The Drake appendix calls Drake the main software toolbox for the text, names Deepnote as the primary course platform, and states that Drake no longer supports Google Colab. Local installation is pip install underactuated[all]. The repository code is BSD 3-Clause; the licence file does not address the book text. [11] [12]

Robotic Manipulation: 12 chapters, updated through Fall 2026

Robotic Manipulation: Perception, Planning, and Control carries a 2020 to 2026 copyright and says it will be updated throughout the Fall 2026 semester. The Fall 2026 course, 6.4210 and 6.4212, runs from September to December 2026 with Tomas Lozano-Perez, Leslie Kaelbling and Russ Tedrake as instructors. The course page states that all of the course materials are freely available and that assignments use Drake. An archived Fall 2022 edition is on MIT OpenCourseWare. [13] [14] [15]

The twelve chapters are an introduction, choosing a robot, basic pick and place, geometric pose estimation, bin picking, motion planning, mobile manipulation, manipulator control, object detection and segmentation, deep perception for manipulation, reinforcement learning, and soft robots and tactile sensing. Appendices cover spatial algebra, Drake, DrakeGym environments, building a manipulation station and miscellany. [13]

Its Drake appendix still provides Open in Colab links alongside the local route, pip install manipulation[all] with JupyterLab, and visualisation through Meshcat. The repository is BSD-3-Clause. [16] [17]

Drake: what the toolbox provides

Drake describes itself as a collection of tools for analyzing the dynamics of robots and building control systems for them, with a heavy emphasis on optimization-based design and analysis. It is a C++ and Python toolbox with three listed components: multibody kinematics and dynamics, a framework for modelling dynamical systems in a Simulink-inspired diagram style, and MathematicalProgram for solving optimization problems. Integrations with LCM and Python are supported; ROS 2 and Julia integrations are listed as unsupported. [18]

The project began in the Robot Locomotion Group at MIT CSAIL and states that core development is led by the Toyota Research Institute. All components are BSD 3-Clause. Stable releases arrive approximately once a month: v1.54.0 on 12 June 2026, v1.55.0 on 15 July, v1.56.0 on 14 August, v1.57.0 on 10 September and v1.58.0 on 15 October 2026. The October release adds Xcode 27 support, raises the minimum macOS to Tahoe and switches the default binder to nanobind for wheel builds. [18] [19] [20]

Supported platforms are Ubuntu 24.04 and 26.04 LTS and macOS Tahoe on arm64, with Python 3.12 to 3.14 depending on the route, through pip, APT, binary archives or source. The solvers module lists Gurobi, MOSEK, CLP, CSDP, Clarabel, SCS, OSQP, DAQP, SNOPT, Ipopt and NLopt; binary releases include a precompiled SNOPT. The Meshcat class runs an HTTP and WebSocket server for browser visualisation, which is why both textbooks open a Meshcat window. [21] [22] [16]

Papers to read, with their stated numbers

Diffusion Policy, arXiv 2303.04137, evaluates a diffusion model that outputs robot actions on 12 tasks from 4 manipulation benchmarks and reports an average improvement of 46.9 percent over existing methods. Tedrake is on the author list of the arXiv v5 and the IJRR journal version published online on 11 October 2024, as the seventh of eight authors. The RSS 2023 proceedings entry lists seven authors without him, so cite the journal version when attributing the paper to him. [23] [24] [25]

A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation, arXiv 2507.05331 of 7 July 2025, is the TRI LBM paper with Tedrake as last author. It reports pretraining on approximately 1,700 hours of robot demonstrations, over 500 internally collected tasks, a bimanual station of two Franka FR3 arms with parallel grippers, and 1,800 controlled real-world trials evaluated through blind, randomized protocols, with 50 real and 200 simulated rollouts per task, policy and condition. Its stated findings are that finetuning a pretrained LBM needs less than 30 percent of the data required to train from scratch and that performance increases predictably with pretraining scale and diversity. No venue is listed. [26]

Ambient Diffusion Policy: Imitation Learning from Suboptimal Data in Robotics, arXiv 2606.12365, was submitted on 10 June 2026 with Tedrake as last author. The group page records best-paper awards at two RSS 2026 workshops in July 2026. His 28 May 2026 Robotics Summit keynote was titled Building Large Behavior Models for Industry, listed under a stealth startup before the Walden announcement. [27] [3] [28]

• Start with Underactuated Robotics chapters 1 to 5 for the dynamics vocabulary used across his papers.
• Read Robotic Manipulation chapter 3 before Diffusion Policy; the pick-and-place framing is the same.
• Read the LBM paper's evaluation section before its headline numbers; the blind, randomized protocol is what the paper is careful about.

Honors on his own page, and claims left out

His homepage lists the 2024 MIT School of Engineering Distinguished Educator Award, the 2024 MIT EECS Digital Innovation Award, the 2023 MIT Teaching with Digital Technology Award, the 2021 Jamieson Teaching Award, an NSF CAREER Award, the MIT Jerome Saltzer Award, a DARPA Young Faculty Award in Mathematics, the 2012 Ruth and Joel Spira Teaching Award, a Microsoft Research New Faculty Fellowship, and best-paper awards at venues including ICML, RSS, ICRA, CoRL, Humanoids, CVPR, IJRR, WAFR, RA-L, HSCC, ISER and T-RO. The group page adds the ICML 2022 Outstanding Paper award for Do differentiable simulators give better policy gradients? [1] [3]