Chapter 3
Massachusetts Institute of Technology
27 sections · about 24 minutes
3.0 How MIT robotics is organised
MIT's robotics is spread across CSAIL, Mechanical Engineering, AeroAstro, the Media Lab, LIDS and BCS, with no single centre. The nearest thing to an umbrella is a loose federation of three structures: the CSAIL Embodied Intelligence Community of Research, which notably publishes neither a faculty lead list nor a budget; the Embodied Intelligence Seminar Series; and the MIT Quest / Siegel Family Embodied Intelligence mission, whose named leads are an unusual robotics-plus-cognitive-neuroscience pairing: Leslie Kaelbling, Tomás Lozano-Pérez, Nicholas Roy, Joshua Tenenbaum, Nancy Kanwisher and James DiCarlo. The public-facing event is the annual CSAIL Embodied Intelligence Summit, whose second edition in October 2025 was held in Palo Alto and hosted by a venture firm — which tells you where the money and agenda sit.
Two structural facts dominate MIT robotics in 2026, and a textbook should state both plainly.
First, MIT is the intellectual home of the anti-scaling position. At the ICRA 2025 data-versus-models debate, Leslie Kaelbling argued that "data can show us patterns, but models give us understanding," while Daniela Rus took the data-first side. That debate is not a curiosity; it is the field's central disagreement, and MIT hosts both poles of it more sharply than anywhere else.
Second, MIT has been losing robotics people to industry faster than it replaces them. Alberto Rodriguez left for Boston Dynamics and MCube is closing. Sangbae Kim is on leave. Russ Tedrake is simultaneously an MIT professor, TRI's robotics head, and now CEO of Walden Robotics. Julie Shah runs a department. The 2025–26 Schwarzman College of Computing faculty cohort contains no robotics or embodied-AI hires at all. The most consequential recent appointment, Kaiming He (2024), is a vision researcher who also works part-time at Google DeepMind.
3.1 Robot Locomotion Group — Russ Tedrake
Tedrake's group is the only lab in this book running a serious formal-methods programme and an industrial-scale behaviour-cloning programme simultaneously. He is Toyota Professor at MIT, SVP for Large Behavior Models at TRI, and since July 2026 CEO of Walden Robotics, a TRI spinout that launched at a $1.1B valuation with $300M raised, co-led by Toyota and Deviation Capital, building wheeled-base humanoids.
Two research lines. The learning line runs from Diffusion Policy (RSS 2023; IJRR 2025) through the Large Behavior Models programme to the Boston Dynamics × TRI Atlas demonstration (August 2025), plus careful empirical work on sim-and-real co-training of diffusion policies. The rigour line is Graphs of Convex Sets: Shortest Paths in Graphs of Convex Sets (SIAM J. Optimization 2024) and non-Euclidean motion planning with geodesically convex sets (IJRR 2025), plus the Drake toolbox and the Underactuated Robotics course as public infrastructure.
Data stance: unapologetically teleoperation-first — large human demonstration fleets, with simulation as a supplement.
Why promising: Tedrake's group publishes negative and ambiguous results, which almost nobody else in this field does.
Skeptic's view: the LBM results depend on TRI's teleoperation fleet and are effectively unreproducible in academia; and the elegant convex-optimisation work still demonstrates mostly on planning benchmarks rather than contact-rich autonomy in the wild.
3.2 Improbable AI Lab — Pulkit Agrawal
Agrawal's lab holds that dexterity is a force problem, not a position problem, and that the route there is massive simulation plus automatic environment and reward shaping — with better hands and wrists to make the physics learnable at all. The framing is summarised in a December 2025 CSAIL spotlight on force-centric manipulation.
Recent work attacks both ends of the stack. DexWrist (2025) is a compliant wrist for constrained and dynamic manipulation. DEXOP (2025, with Ted Adelson) is a passive hand exoskeleton that mechanically couples a human's fingers to a robot hand's kinematics while capturing full-hand tactile data — "perioperation," and one of the most important data-collection ideas of the last two years. Alongside: Bridging the Sim-to-Real Gap for Athletic Loco-Manipulation (RSS 2025), Robot Learning with Super-Linear Scaling (RSS 2025), DexHub and DART (ICRA 2025), Vegetable Peeling (ICRA 2025, with TRI), and the position paper Automatic Environment Shaping is the Next Frontier in RL (ICML 2024 oral).
(Correction worth flagging: DrEureka is a UPenn/NVIDIA paper, not an Improbable AI product. The MIT analogue is the automatic-environment-shaping and reward-selection line.)
Why promising: it attacks sensors, mechanisms and algorithms rather than only policy architectures — and DEXOP may be the cheapest route to a tactile-rich manipulation corpus.
Skeptic's view: results remain largely per-task RL with heavy shaping; "super-linear scaling" rests on curated task suites; and no Improbable AI policy yet runs unattended in an uninstrumented home.
3.3 Learning and Intelligent Systems — Leslie Pack Kaelbling & Tomás Lozano-Pérez
LIS is the field's most articulate anti-scaling group. Its thesis: general-purpose robots need abstraction — hierarchical, hybrid discrete-continuous task and motion planning (TAMP), with learning supplying the samplers, predicates and skill parameters rather than the whole policy.
The 2025–26 work is a genuinely new development in an old programme, because VLMs turn out to solve TAMP's historic bottleneck — where do the symbols come from? From Pixels to Predicates: Learning Symbolic World Models via Pretrained VLMs (RA-L 2026) and Open-World TAMP via VLM-Generated Constraints (RA-L 2026) have a foundation model invent the predicates and constraints that a planner then reasons over. Around them: TiPToP, a modular open-vocabulary planning system (2026); Differentiable GPU-Parallelized TAMP (RSS 2025); Streaming Flow Policy (CoRL 2025, with Nicholas Roy); Practice Makes Perfect: Planning to Learn Skill Parameter Policies (RSS 2024, with the Boston Dynamics AI Institute); and Trust the PRoC3S (CoRL 2024).
Data stance: explicitly anti-scaling; heavy use of simulation; evidence is more often simulated than physical.
Why promising: VLM-generated predicates are a real answer to the symbol-grounding problem that stalled TAMP for thirty years, and TAMP remains the only framework with principled guarantees on long-horizon tasks.
Skeptic's view: demonstrations stay close to tabletop and blocks-world domains; invented predicates inherit VLM hallucination; and the group's real-robot footage is thin relative to its paper output. It is also worth noting that Gemini Robotics' "think before acting" is arguably TAMP re-implemented in natural language — which either vindicates the programme or makes it redundant.
3.4 Distributed Robotics Lab — Daniela Rus
Rus, CSAIL's director, holds that intelligence should be distributed across body and controller — soft and modular morphology plus compact continuous-time networks, rather than large monolithic policies.
The standout recent result is Controlling diverse robots by inferring Jacobian fields with deep networks (Nature, June 2025, with Vincent Sitzmann and Sizhe Lester Li): robots learn to control themselves from monocular video alone, with no kinematic model. In January 2026 came a neural blueprint for soft-robot control using "structural" and "plastic" synapses (Science Advances, with NUS), holding above 92% shape accuracy under 50% actuator failure. Also: RoboGrocery (RoboSoft 2024), tactile soft grippers packing groceries; AI-designed underwater gliders (ICRA 2025, with Matusik); and GPU convex-set computation (2025, with Tedrake). The liquid-network line spun out as Liquid AI.
Data stance: pro-data — she argued the data-first position at the ICRA 2025 debate.
Why promising: the only MIT lab systematically pairing novel morphology with novel network classes; Neural Jacobian Fields is a genuinely surprising result about how little structure a controller needs.
Skeptic's view: enormous breadth and demo-shaped results. Neural Jacobian Fields still needs multi-camera retraining per robot and has no force sensing; liquid networks' efficiency claims remain contested outside small control benchmarks.
3.5 MCube Lab — Alberto Rodriguez (closing)
Status: effectively defunct. The MCube site states that Rodriguez "has stepped down from MIT and transitioned full time to a new adventure at Boston Dynamics. The MCube lab will close in the near future." He is now Director, Robot Behavior – Atlas.
The legacy thesis: manipulation is fundamentally about contact mechanics. MCube produced the canonical planar-pushing datasets, the extrinsic dexterity framework (using walls, gravity and the environment as extra fingers), tactile-informed regrasping, GelSlim tactile fingers (with Adelson), and the MIT-Princeton Amazon Robotics Challenge picking system. Much of the current contact-rich learning literature still benchmarks against MCube's empirical mechanics.
Why it still matters: the contact-rich agenda did not die, it migrated. Rodriguez now leads behaviour for Atlas — the same platform running Tedrake's TRI Large Behavior Models — so MCube's intellectual lineage sits inside the highest-profile humanoid programme in the US.
Skeptic's view: this is the clearest case study in MIT robotics' talent drain. A decade of rigorous contact mechanics no longer has an academic home, and the successor work is behind corporate walls where reproducibility is impossible.
3.6 GelSight / Perceptual Science Group — Edward Adelson
Adelson solved touch as a vision problem: image an elastomer gel from the inside with a camera and you recover micron-scale geometry, shear and force from commodity parts. GelSight is the most widely adopted tactile sensing technology in robotics, and essentially every optical tactile sensor since is a descendant.
Recent work: Object Recognition and Force Estimation with the GelSight Baby Fin Ray (2025), classifying in-shell nuts by touch while digging through clutter; DEXOP (2025, with Agrawal), the tactile-instrumented hand exoskeleton; building on the platform papers GelSight Svelte (2023) and GelSight Fin Ray (2022). GelSight Inc. commercialises the metrology side, and Meta's Digit 360 is a direct descendant.
Data stance: hardware-first. Build a sensor good enough that modest datasets suffice; no interest in scaling teleoperation for its own sake.
Why promising: DEXOP suggests the field's real bottleneck may be data collection ergonomics, and tactile-rich human demonstration is the most plausible answer anyone has offered.
Skeptic's view: GelSight's problems are unchanged since 2017 — gel wear, recalibration, camera bulk, latency — and the research centre of gravity has moved to spinouts, Wenzhen Yuan's group, and Meta. It also remains unproven that touch beats good vision plus mechanical compliance for most manipulation tasks.
3.7 Biomimetic Robotics Lab — Sangbae Kim (dormant)
Status: on leave. MIT MechE lists Kim as a professor "(On Leave)"; trade press widely reports him as Robotics Architect at Meta working on its humanoid programme, though this is not confirmed by any MIT primary source. The lab is not actively led.
The thesis: biology gives you mechanism, not algorithms. High-bandwidth, backdrivable proprioceptive actuators — the "MIT actuator" — make dynamic legged behaviour a control problem rather than a sensing problem. The lab produced Mini Cheetah (first quadruped backflip), Cheetah 3, and the MIT Humanoid. Recent publications include CusADi (GPU parallelisation of symbolic expressions for optimal control, RA-L 2024), FLD: Fourier Latent Dynamics (2024), integrating model-based footstep planning with model-free RL (IROS 2024), and URDF+ extending URDF to kinematic loops (Humanoids 2024). No 2025–26 publications appear on the lab site — possibly a stale page rather than a true gap.
Data stance: the most model-based group here — whole-body MPC first, RL as a bolt-on. Kim has been publicly sceptical of end-to-end learning for legged systems.
Why promising: the actuator design philosophy is now industry standard, and Kim's move to Meta is a strong signal about where humanoid hardware credibility sits.
Skeptic's view: Unitree and other Chinese manufacturers have commoditised quasi-direct-drive actuation, eroding the original moat, and the lab has no active successor.
3.8 Interactive Robotics Group — Julie Shah
Shah — head of AeroAstro since 2024 — works on robots as teammates: mutual prediction, shared mental models, calibrated trust and legible explanation, rather than raw dexterity.
The single most relevant recent paper for the VLA era is Inference-Time Policy Steering Through Human Interactions (ICRA 2025, with NVIDIA), which lets a human nudge a diffusion policy at inference time — exactly the interface layer that generative manipulation policies currently lack. Around it: REALM, real-time estimates of assistance for learned models (RA-L 2025); a Versatile Demonstration Interface (IROS 2025); Questioning the Robot: Using Human Non-verbal Cues to Estimate the Need for Explanations (HRI 2025); Learning Contextually-Adaptive Rewards via Calibrated Features (HRI 2026, with Andreea Bobu); Adaptive Language-Guided Abstraction from Contrastive Explanations (CoRL 2024). Historically the group deployed on BMW assembly lines and in hospital labour-and-delivery scheduling.
Data stance: human-in-the-loop demonstration and correction, deliberately low-N, treating human factors evidence as first-class data.
Skeptic's view: the group's public site has not been updated since roughly 2023, department-head duties consume bandwidth, and HRI's small-sample user studies replicate unevenly. The group is a consumer rather than a producer of the foundation-model wave.
3.9 Computational Design and Fabrication Group — Wojciech Matusik
CDFG argues that morphology should not be fixed: co-optimise structure and controller using differentiable simulation, then actually fabricate the result.
Recent work: AI-enhanced automatic design of efficient underwater gliders (ICRA 2025, with Rus); Learning Object Properties Using Robot Proprioception via Differentiable Robot-Object Interaction (ICRA 2025); Adaptive Walker, an intent- and terrain-aware walker with high-resolution tactile and IMU sensing (ICRA 2025); a modular self-reconfigurable continuum robot for loco-manipulation (RA-L 2025); electronic-free particle robots communicating through architected tentacles (Advanced Intelligent Systems 2025); the WiReSens Toolkit for accessible wireless tactile sensing (TEI 2026); and ASAP assembly-sequence planning (ICRA 2024).
Why promising: machine-knitted, cheap, distributed tactile sensing is one of very few plausible routes to whole-body touch at scale — which is the modality Chapter 13 identifies as most conspicuously missing.
Skeptic's view: co-design results overwhelmingly stay in simulation or in single bespoke prototypes, and SIGGRAPH-inflected contributions rarely appear on manipulation benchmarks.
3.10 SPARK Lab — Luca Carlone
SPARK argues perception needs certificates — algorithms with provable global optimality or verifiable failure detection — plus hierarchical, task-aware representations that go beyond metric SLAM.
The code ships and is widely used, which distinguishes this lab: Hydra builds real-time 3D scene graphs; Clio makes them task-driven and open-set; Khronos (RSS 2024) does spatio-temporal metric-semantic SLAM in dynamic environments; VGGT-SLAM (NeurIPS 2025) does dense RGB SLAM on the SL(4) manifold; BUFFER-X (ICCV 2025 Highlight) does zero-shot point-cloud registration; Kimera-Multi supplies multi-robot SLAM datasets. Carlone is currently on sabbatical as an Amazon Scholar.
Why promising: 3D scene graphs are a credible interface between geometry and LLM planners — arguably the missing representation between "the robot sees pixels" and "the robot reasons about rooms."
Skeptic's view: certifiability results mostly cover convex relaxations of narrow subproblems, not the learned front-ends now doing the heavy lifting. The guarantee sits on the wrong side of the pipeline.
3.11 Robust Robotics Group — Nicholas Roy
Roy's group works on autonomy under uncertainty with no prior map, combining decision-theoretic planning, statistical inference and AI.
The distinctive 2026 contribution is a genuine third way on foundation models: Belief Consistency Between Foundation-Model Evidence and Geometric Perception in Persistent Robotic Maps (2026) treats VLM outputs as evidence to be reconciled with geometry, not as ground truth. Around it: Belief Roadmaps with Uncertain Landmark Evanescence (ICRA 2025); Anomalies-by-Synthesis, diffusion-based anomaly detection for off-road navigation (ICRA 2025); PIETRA, physics-informed evidential learning for out-of-distribution terrain (RA-L 2025); FORGE, force-guided exploration for contact-rich manipulation under uncertainty (RA-L 2025); language-grounded hierarchical planning with multi-robot 3D scene graphs (2025); and POMDP learning theory beyond full-rank actions (2026).
Why promising: as VLA policies proliferate, someone must formalise when to trust them, and this group has the POMDP machinery.
Skeptic's view: the portfolio is diffuse and much output lands in RA-L and preprints rather than agenda-setting venues.
3.12 Marine Robotics Group — John Leonard
Leonard's group works on durable long-term mapping: consistent spatial representations across months and changing environments, with underwater as the hardest case.
Recent work: ReefMapGS, closing the loop between multimodal SLAM and Gaussian splatting for large-scale underwater reconstruction (2026); SeaSplat, 3D Gaussian splatting for underwater scenes (ICRA 2025); NeuSE, SE(3)-equivariant embeddings for long-term object-based SLAM (IJRR 2026); 3DGS-CD, splatting-based change detection for object rearrangement (RA-L 2025); opti-acoustic semantic SLAM with unknown objects (IROS 2024); semantic enhancement for object SLAM using multimodal LLM agents (IROS 2025).
Why promising: underwater reconstruction has real scientific demand (reef monitoring) and no plausible internet-scale data shortcut — a useful counterexample to the "everything is a data problem" consensus.
Skeptic's view: the group's centre of gravity has drifted toward whatever is fashionable in 3D vision; the distinctive marine contribution can look like generic SLAM plus salt water.
3.13 Aerospace Controls Lab — Jonathan How
ACL works on decision-making under uncertainty for vehicle teams: distributed planning, robust and safe control, and learned policies with formal guarantees.
Recent: GRAND-SLAM, globally consistent large-scale multi-agent Gaussian SLAM (2025); LunarLoc, global localisation for lunar surface operations (2026); SANDO and MIGHTY for safe trajectory planning in dynamic unknown environments (2026); PIETRA (with Roy); aerobatic manoeuvres in insect-scale flapping-wing robots via deep-learned robust tube MPC (2025); CLIPPER, robust data association without an initial guess (2024); GUARDIAN, safety filtering for perception models under adversarial attack (2026).
Why promising: DoD- and NASA-relevant, verification-oriented, and strong on multi-robot SLAM through the ARL DCIST alliance.
Skeptic's view: enormous publication volume with incremental acronym churn, and safety guarantees typically assume perception models the same papers elsewhere admit are the weak link.
3.14 d'Arbeloff Laboratory — Harry Asada
Asada's lab builds robots that are worn, braced against, or fused with the human body — supernumerary hardware that adds degrees of freedom to a person — plus contact-rich factory automation.
Recent: E-BAR (ICRA 2025), a mobile eldercare robot with inflatable airbags that supports sit-to-stand and physically catches falls; Loop Closure Grasping (Science Advances 2025), topological grasp transformations for strong-yet-gentle grips; soft growing structures for patient transfer (IROS 2025); Koopman dynamic modelling of rigid bodies with contact (IROS 2024); supernumerary robotic limbs including a spacesuit variant.
Why promising: eldercare and aircraft assembly are real markets with hard physical constraints that foundation models do not address, and Loop Closure Grasping is a genuinely novel mechanism idea.
Skeptic's view: a long history of striking demos with thin longitudinal human-subject validation; supernumerary limbs have been "nearly deployable" for a decade.
3.15 Media Lab Personal Robots Group — Cynthia Breazeal
Breazeal founded social robotics as a field; she is now also MIT's Dean for Digital Learning, and her personal recent output is dominated by K-12 AI literacy rather than robots. The group remains active, with research scientists carrying much of the robotics.
The standout result is Social robots as conversational catalysts: enhancing long-term human-human interaction at home (Science Robotics, March 2025, front cover) — a robot that improves interaction between people rather than with itself. Also: Social Robots as Social Proxies (2025), a 40-home two-week deployment on empathy; the EMPATHICSTORIES++ dataset; Doodlebot; Tega for early literacy; an HRI 2024 best paper on adaptive robot roles.
Why promising: genuinely long-horizon, in-home studies — rare, hard, and exactly the data the rest of the field lacks.
Skeptic's view: the commercial base collapsed with Jibo, effect sizes in wellbeing studies are modest, and LLM chat now delivers much of the perceived value without a robot.
3.16 Center for Bits and Atoms — Neil Gershenfeld
CBA proposes replacing continuous fabrication with digital materials: reversible, error-correcting assembly of discrete voxels by simple robots that crawl on the structure they build.
The 2026 result is Comparative evaluation of robotically assembled discrete lattice systems for sustainable construction (Automation in Construction, April 2026), introducing MILAbots (Modular Inchworm Lattice Assemblers) and octet-lattice voxels in steel and wood, with a one-storey demonstrator, modelling of up to 20 robots in parallel, and up to 82% embodied-carbon reduction versus 3D concrete printing. Earlier: self-reconfigurable robots for collaborative discrete lattice assembly (ICRA 2024).
Why promising: decouples robot complexity from structure size, and builds disassembly and reuse in architecturally, which matters for construction carbon.
Skeptic's view: three decades of self-replication rhetoric with no structural-engineering certification, no cost parity, and lab-scale demonstrations. The 2026 paper's assembly rates at scale are modelled, not measured.
3.17 K. Lisa Yang Center for Bionics — Hugh Herr
Herr argues the limitation in prosthetics is not the actuator but the interface: surgically reconstruct the residual limb to restore proprioception, then let the person's own nervous system control the machine.
The landmark result is Tissue-integrated bionic knee restores versatile legged movement after amputation (Shu et al., Science 389:eadv3223, July 2025), combining osseointegration, permanently implanted hardware and neuromuscular signals from surgically modified tissue — building on the agonist-antagonist myoneural interface (AMI) procedure and the 2024 Nature Medicine fully neural-controlled bionic leg.
Embodiment relevance: this is the strongest evidence anywhere that embodiment is a bidirectional problem — the body must be redesigned for the machine, not only the machine for the body.
Skeptic's view: AMI requires elective surgical modification, cohorts are single-digit, and cost puts it decades away from most amputees. The rhetoric outruns the clinical evidence.
3.18 GEAR Center — Amos Winter
GEAR derives device requirements from first-principles biomechanics so that passive, low-cost, mass-manufacturable prostheses match the gait quality of expensive powered ones.
The core method is the Lower Leg Trajectory Error (LLTE) metric, mapping mechanical design directly to biomechanical outcome, extended in 2025 to a Hip Trajectory Error framework for above-knee amputation (Journal of Mechanical Design). Products include a single-part injection-moulded prosthetic foot manufactured in India with month-long clinical trials, and a low-cost passive knee. Partners include BMVSS/Jaipur Foot and Northwestern's Prosthetics-Orthotics Center.
(Correction: Faye Wu is not at GEAR. She is an Asada PhD, now co-founder/CTO of Manus Robotics, building multimodal muscle-activity sensing for assistive-device control.)
Why promising: rigorous design theory tied to a distribution channel already serving millions.
Skeptic's view: essentially no manipulation and no autonomy — optimised mechanism design, belonging more to an assistive-devices chapter than a robotics one.
3.19 Kaiming He — the vision substrate
He joined MIT EECS in 2024 and has been a part-time Distinguished Scientist at Google DeepMind since 2025. He is not a roboticist; he builds the representation-learning substrate embodied systems consume (ResNet, Mask R-CNN, MoCo, MAE).
His one genuinely embodied paper is important: Heterogeneous Pre-trained Transformers (HPT) (Wang, Chen, Zhao, He; NeurIPS 2024), which pre-trains a shared policy "trunk" across robot embodiments with embodiment-specific "stems" for heterogeneous proprioception and vision, and task-specific "heads." Over 50 datasets and 200k+ trajectories it beat baselines by more than 20% on unseen tasks. It is the cleanest MIT statement of the "robot data is heterogeneous, pre-train the trunk" thesis. His generative work — Mean Flows for One-step Generative Modeling (NeurIPS 2025 Oral), Is Noise Conditioning Necessary for Denoising Generative Models? (ICML 2025) — attacks diffusion-policy inference latency directly.
Skeptic's view: treating robotics as a downstream application of vision has repeatedly under-delivered on contact and dynamics, and the DeepMind affiliation means his highest-impact work may not land at MIT.
3.20 Phillip Isola's group
Isola argues that intelligence is representation, and that models trained on different modalities converge toward a shared statistical model of reality — The Platonic Representation Hypothesis (ICML 2024). For robotics the corollary is: if generative models already encode the visual world, use them to manufacture the data robots need.
LucidSim (CoRL 2024, with Alan Yu, Ge Yang and John Leonard) is the existence proof: a Unitree Go1 learned vision-based parkour entirely from generative-model-rendered scenes, with zero real training images. LucidXR (CoRL 2025) extends this to dexterous manipulation from human demonstrations captured in XR.
Data stance: the purest "generative simulation" position at MIT — synthesise photorealistic variation to close the appearance half of the sim-to-real gap, while physics stays in a conventional simulator.
Skeptic's view: generated images fix textures, not contact dynamics, friction or deformables — precisely where manipulation breaks. And the Platonic hypothesis is a position paper with contested evidence.
3.21 Antonio Torralba's group and MIT-IBM Watson AI Lab
Torralba, Faculty Head of AI+Decision Making, works on multimodal representation, interpretability and simulated worlds as substrate for embodied learning. Recent: Separating Knowledge and Perception with Procedural Data (ICML 2025); Adaptive Length Image Tokenization via Recurrent Allocation (ICLR 2025); Eval3D and SketchAgent (CVPR 2025); A Multimodal Automated Interpretability Agent (ICML 2024); the Foundations of Computer Vision textbook (2024). Earlier embodied infrastructure includes the ThreeDWorld simulation platform.
Honest framing: this is adjacency, not embodiment. No substantial first-party robot-hardware papers were found for 2025–26, and robotics influence flows indirectly — through alumni (Yunzhu Li, Chuang Gan), simulators, and interpretability tools that robot-learning groups adopt. Its inclusion is justified as infrastructure and lineage, not as a robotics lab.
3.22 Jeehwan Kim — neuromorphic hardware (robotics relevance: weak)
Kim's lab works on analogue neuromorphic hardware — dense artificial-synapse arrays, 2D-material and remote epitaxy, 3D heterogeneous integration — so inference can run at the edge with far lower energy. Named work: tens of thousands of artificial brain synapses on a single chip (2020); a "LEGO-like" stackable AI chip (2022); 3D integration of AI hardware with direct analogue input from sensor arrays under the MIT AI Hardware Program.
Explicit flag: do not present this as a robotics group. No robotics-specific project, robot platform, or robotics publication was found for 2024–26. The honest framing is a materials and device group whose sensor-adjacent analogue compute is a plausible future substrate for low-power embodied inference, with no robot currently in the loop. If analogue in-memory compute matures, on-robot perception power budgets change qualitatively; neuromorphic hardware has also repeatedly failed to displace GPUs.
3.23 Toyota Research Institute — the MIT-industry axis
TRI's robotics division is led by Russ Tedrake under a joint appointment. The flagship result is A careful examination of large behavior models for multitask dexterous manipulation (Barreiros, Beaulieu, Tedrake et al., Science Robotics 11(113), April 2026) — a 94-author paper using blind evaluators, 1,800 real-world trials and 50 rollouts per task-policy pair, which found that diffusion-based multitask pretraining genuinely improves robustness and few-shot acquisition, requiring "less than 30% of the data needed for training from scratch," but with smooth, gradual gains and no discontinuity at current scales.
Precursors: TRI's 2023 diffusion-policy work and the Boston Dynamics–TRI Atlas partnership (October 2024). Recent joint MIT-TRI work includes SceneSmith (ICML 2026 spotlight), agentic generation of simulation-ready indoor scenes.
Why it matters: this is the most statistically careful evaluation in a field of cherry-picked videos, and its conclusions are deliberately unflattering to its own programme.
Skeptic's view: the paper's own finding — real but modest, data-hungry gains — is a long way from the general-purpose claims made around it. The MIT/TRI/Walden boundary is now blurred enough to complicate attribution entirely.
3.24 MIT spinouts: an attribution audit
Verified MIT lineage. Walden Robotics (Cambridge, 2026) — Tedrake as CEO, TRI spinout, $300M at $1.1B, wheeled-base humanoids trained by large behaviour models, already running in a Toyota plant. Liquid AI (Hasani, Lechner, Amini, Rus) — a direct CSAIL spinout from liquid neural networks, though now focused on efficient LLMs rather than robots. Manus Robotics (Asada, Faye Wu). RightHand Robotics — partial lineage; co-founder Lael Odhner did his degrees with Asada and came out of the 2009 DARPA ARM programme, though other co-founders are Harvard. Historic: Boston Dynamics (Marc Raibert, MIT Leg Lab) and iRobot (Brooks, Angle, Greiner). The RAI Institute (Raibert, Cambridge) has absorbed considerable local talent.
Corrections to flag. Realtime Robotics is a Duke spinout, not MIT. Ambi Robotics is a UC Berkeley (Ken Goldberg) spinout. Do not attribute either to MIT.
Skeptic's view: a $1.1B valuation six months from founding and pre-product is the clearest single sign of a humanoid funding bubble, and a sitting MIT professor as startup CEO raises the conflict questions that dogged the last two robotics cycles.
3.25 The hiring picture
There was no significant MIT robot-learning faculty hire in 2025–26 and no new MIT robotics centre. The 2025–26 Schwarzman College of Computing cohort — eleven people including Omar Khattab (NLP/IR) and Mitchell Gordon (HCI+ML) — contains no robotics or embodied-AI appointments. The 2024–25 School of Engineering cohort yielded one adjacent hire, Raphael Zufferey (MechE), on bio-inspired aerial-aquatic locomotion.
The substantive institutional change is instead the MIT Initiative for New Manufacturing (May 2025), Institute-wide with an AI-and-automation framing.
Skeptic's reading: MIT is losing the robot-learning hiring race. The field's rising stars went to Stanford, Berkeley, CMU and Columbia, while MIT's marginal robotics capacity flowed out to TRI, Walden, Boston Dynamics, Meta and RAI. A textbook should say so.
3.26 Deep dive: MIT's real contribution is the objection
Stanford's contribution to 2023–2026 was interfaces. MIT's was the objection — and the field is better for it.
Four objections, each with a serious lab behind it.
The evaluation objection (Tedrake/TRI). Before the Large Behavior Models paper, the modal robot-learning result was a video and a success rate from an unspecified number of trials scored by the authors. TRI's methodology — blind evaluators, 1,800 trials, 50 rollouts per task-policy pair, 27% re-scored for quality control, with a measured 2.31% inter-grader discrepancy — established that much prior reporting was too noisy to support its own claims. It is difficult to overstate how much this should change how the field reads its own literature.
The abstraction objection (Kaelbling/Lozano-Pérez). Long-horizon manipulation is a hybrid discrete-continuous search problem, and end-to-end policies do not have the structure to solve it. The 2026 twist is that this group is now using foundation models to supply the symbols rather than fighting them — a much stronger position than the pure-symbolic one it replaced.
The tactile objection (Adelson, Agrawal, and Rodney Brooks from outside). Brooks' September 2025 essay puts it most sharply: "Collecting just visual data is not collecting the right data," because human dexterity runs on roughly 17,000 hand mechanoreceptors, and "we as a species have not developed technologies to capture touch, to store touch, to transmit touch over distances and time." DEXOP is the constructive answer — a passive exoskeleton that captures full-hand tactile data in the robot's own action space — and it is arguably MIT's most important 2025 artifact.
The mechanism objection (Kim, Cutkosky from Stanford, Herr, Asada). Some problems are solved by better bodies, not better policies. Proprioceptive actuators made dynamic legged locomotion tractable. Loop closure grasping makes a grip strong and gentle at once. Osseointegration plus AMI restores proprioception that no controller could synthesise. This objection is unfashionable and probably correct at the margin.
The tension worth noticing is that all four objections are, in 2026, being absorbed rather than defeated. Gemini Robotics thinks before acting (abstraction). Figure and Apptronik ship fingertip tactile sensing (touch). Everyone now reports more trials (evaluation). Every humanoid company is a hardware company (mechanism). Whether that absorption vindicates the objectors or dissolves them into the scaling programme is the open question of the next three years.