Chapter 2
Stanford University
18 sections · about 20 minutes
2.0 How Stanford robotics is organised
There is no "Stanford Robotics Department." Robotics is federated across the Stanford AI Lab (directed by Carlos Guestrin since 2025, succeeding Christopher Manning), the CS department, Mechanical Engineering, Aeronautics & Astronautics, EE, and Bioengineering. The Stanford CS robotics roster lists thirteen faculty; the hardware and controls groups sit outside CS entirely.
The integrating layer, since November 2024, is the Stanford Robotics Center, directed by Oussama Khatib, occupying a converted basement of the Packard Electrical Engineering Building. Stanford News reported more than 30 faculty across nine departments at opening; later Center materials cite 45 faculty across ~18 labs, with an executive committee of Cutkosky, Okamura, Pavone, Sadigh and Wu. (Verification note: the Center's own site does not currently state an opening date or name a director; the November 2024 date and Khatib's directorship come from Stanford News and Stanford HAI announcements.)
Stanford's distinctive contribution to the 2023–2026 period is not a single model. It is hardware and interfaces that made data collection cheap enough for everyone else: ALOHA, Mobile ALOHA, UMI, DexUMI, TidyBot++, ToddlerBot, TWIST2. Every one of these was built by graduate students on small budgets and then copied globally. The counterpart weakness is equally clear: the people who built them now largely work at Physical Intelligence, Sunday Robotics, and Google DeepMind.
2.1 Stanford Vision and Learning Lab (SVL) — Fei-Fei Li & Jiajun Wu
SVL is co-directed by Fei-Fei Li (also co-director of Stanford HAI, and since 2024 CEO of World Labs, which has drawn much of her attention toward "spatial intelligence") and Jiajun Wu, with Ruohan Zhang as a research lead. The thesis is that embodied intelligence should be measured against what people actually want robots to do — so build the benchmark and the data engine first, and let the models follow.
The flagship is BEHAVIOR-1K: 1,000 household activities derived from surveys of what people want automated, across 50 interactive scenes and 10,000+ objects, running on OmniGibson, an Isaac Sim-based simulator that models fluids, cloth, heat and transparency. Around it: the BEHAVIOR Robot Suite (CoRL 2025), which introduced the JoyLo whole-body teleoperation interface on a Galaxea R1 wheeled bimanual robot; ReKep (CoRL 2024), which has a VLM generate relational keypoint constraints that a solver then optimises; and the 2026 BEHAVIOR Challenge — 100 full-length household tasks, ~20,000 teleoperated demonstrations totalling 1,950 hours, with π0.5 and GR00T N1.7 as baselines and an October 2026 deadline.
Data stance: sim-first, teleop-heavy. Scale demonstrations inside a high-fidelity simulator, then transfer.
Why promising: it is the only effort anywhere attacking multi-minute, multi-room, state-change household tasks with a shared scoreboard and partial-credit scoring (BDDL goal predicates). Nearly every other benchmark measures single-shot tabletop skills.
Skeptic's view: OmniGibson success may simply not predict real-world success; 1,950 hours of simulated teleoperation is small next to industry real-robot fleets; and fluids, cloth and contact are precisely the physics simulators get most wrong.
2.2 IPRL — Interactive Perception and Robot Learning (Jeannette Bohg)
IPRL argues that manipulation competence comes from interactive perception: the robot must act in order to perceive, and force and contact signals are first-class citizens rather than an afterthought to vision.
Recent work is unusually varied in data source. Masquerade (ICRA 2026) learns from in-the-wild human video by editing the video so the human's arms look like robot arms, closing the visual embodiment gap at the pixel level. HoMeR (ICRA 2026, with Sadigh) does in-the-wild mobile manipulation via hybrid imitation plus whole-body control. DexForce (2025) extracts force-informed actions from kinesthetic demonstrations. Crossing the Human–Robot Embodiment Gap with Sim-to-Real RL Using One Human Demonstration (CoRL 2025, with Karen Liu) does what the title says. SimToolReal (RSS 2026) targets zero-shot dexterous tool use. And TidyBot++ gave the community a widely-copied low-cost holonomic mobile manipulator.
Data stance: the most pluralist lab at Stanford. Human video, kinesthetic teaching, teleoperation and sim-to-real RL are treated as substitutable sources, with data editing used to bridge between them cheaply.
Why promising: human video is the only genuinely abundant manipulation data, and Bohg's group has the most credible recipes for actually using it.
Skeptic's view: video-editing and retargeting tricks discard exactly the contact forces this lab argues are essential. Each paper's pipeline is bespoke; there is no single scaling law on offer.
2.3 ILIAD — Intelligent and Interactive Autonomous Systems (Dorsa Sadigh)
ILIAD treats robots as agents inside a loop with humans — so learn from preferences, corrections, language and interaction, and treat data curation as a first-class algorithmic problem rather than a data-engineering chore. Sadigh has a long-running affiliation with Google DeepMind's robotics effort and was on partial leave through 2025–2027.
She was a principal driver of Open X-Embodiment / RT-X, the 22-embodiment consortium dataset that remains the field's canonical cross-embodiment corpus. Her group's more distinctive line is about which data to collect and keep: RoboCrowd (ICRA 2025) and RoboCade (ICRA 2026) crowdsource and gamify teleoperation so non-experts will produce demonstrations; Robot Data Curation with Mutual Information Estimators (RSS 2025) argues explicitly that more data is not automatically better; A Taxonomy for Evaluating Generalist Robot Manipulation Policies (RA-L 2026) attacks the evaluation problem; Robot-Powered Data Flywheels (2026, with Bohg) closes the loop.
Data stance: real teleoperation at scale, but curated — the interesting question is which demonstrations earn their place.
Why promising: cross-embodiment pooling is the closest robotics has come to an ImageNet moment, and curation is the obvious next lever once raw collection saturates.
Skeptic's view: Open X-Embodiment is a heterogeneous union of many small datasets, and RT-X's gains were modest and largely in-distribution. Crowdsourced teleoperation may buy quantity at the cost of the precision that contact-rich tasks require.
2.4 IRIS — Chelsea Finn
A common error worth correcting: IRIS ("Intelligence through Robotic Interaction at Scale") is Chelsea Finn's group. REAL (Robotics and Embodied AI Lab) is Shuran Song's. They are frequently conflated.
Finn's position is that generalist robot policies come from scaling imitation learning on cheap, open hardware — and she has done more than anyone to make that hardware exist. ALOHA (RSS 2023) introduced both a sub-$20k bimanual teleoperation rig and ACT (Action Chunking with Transformers); Mobile ALOHA (2024) added a wheeled base at ~$32k and became the field's default low-cost platform. Her group co-produced OpenVLA and Octo, two of the most-used open VLAs, and DROID, the standardised in-the-wild manipulation dataset. RoboReward (January 2026, with Levine and Pertsch) builds general-purpose VLM reward models plus a benchmark, attacking the problem of how you score a robot's behaviour without a human watching.
Finn co-founded Physical Intelligence in 2024, which is where much of this agenda now runs at scale.
Data stance: real-world first; minimise the cost per hour of collection rather than substituting simulation.
Why promising: the ALOHA lineage demonstrably transfers — the recipes work in other people's labs, which is rare.
Skeptic's view: the frontier has migrated to startups. The academic lab increasingly does hardware and benchmarks whose payoff is captured elsewhere, and imitation-only policies remain brittle outside their demonstration distribution.
2.5 REAL — Robotics and Embodied AI Lab (Shuran Song)
Shuran Song came to Stanford EE from Columbia (2023; Stanford's own profile confirms the move but not the year). Her thesis: manipulation intelligence comes from scalable, cheap, embodiment-agnostic data collection plus expressive generative policies. Build the interface, not just the algorithm.
The defining artifact is the Universal Manipulation Interface (UMI) (RSS 2024): a 3D-printable handheld parallel-jaw gripper with a wrist-mounted GoPro that lets anyone collect in-the-wild demonstrations without owning a robot, calibration-free, deployable in a new environment in about two minutes. The family has since grown: UMI on Legs (CoRL 2024), DexUMI (CoRL 2025, Best Paper Finalist) extending the interface to multi-fingered hands via a wearable exoskeleton with robot-hand video inpainting, UMI-on-Air (ICRA 2026) for aerial embodiments, and UMI-Underwater. Alongside: ToddlerBot (CoRL 2025), a ~$6k open-source 3D-printed humanoid; Adaptive Compliance Policy (ICRA 2025); DexMachina (ICML 2026). Diffusion Policy itself (Chi, Song, Tedrake et al., RSS 2023) originated in her Columbia group and is arguably the single most influential architecture in modern manipulation.
Data stance: relentlessly low-cost and open. Consumer cameras, printed parts, everything on GitHub.
Why promising: UMI decoupled data collection from robot ownership, which is the field's biggest structural bottleneck, and it has been replicated worldwide.
Skeptic's view: UMI data has no force or proprioceptive channel and suffers camera-pose drift; "in-the-wild" collection still requires disciplined operators; and the scaling laws for this data class remain unproven.
2.6 The Movement Lab — C. Karen Liu
The Movement Lab holds that physics-based character animation and humanoid robotics are the same problem, and that simulation of contact-rich human movement is the bridge between them. In a March 2026 Stanford interview Liu framed the goal as a pretrained humanoid foundation model roughly as capable as a ten-year-old, and defended academia's distinct role: "Industry tends to quickly converge on a recipe that works, and then scale up as fast as possible."
The lab's most consequential recent output is TWIST (CoRL 2025) and TWIST2 (ICRA 2026), a teleoperated whole-body humanoid imitation system that became a portable, mocap-free data-collection rig — VR headset plus a $250 two-degree-of-freedom robot neck, claiming 100 demonstrations in fifteen minutes. Around it: Hand-Eye Autonomous Delivery (CoRL 2025) fusing humanoid navigation, locomotion and reaching; Learning to Ball (SIGGRAPH Asia 2025) composing long-horizon basketball policies; Human-Object Interaction from Human-Level Instructions (ICCV 2025); PGC: Physics-Based Gaussian Cloth (CVPR 2025); and Learning Humanoid Navigation from Human Data (RA-L 2026, with Monroe Kennedy).
Data stance: human motion capture and video, retargeted onto humanoids, trained in simulation, augmented with whole-body teleoperation.
Why promising: nobody has better priors on the physics of human contact, and whole-body humanoid data collection is the field's newest bottleneck.
Skeptic's view: retargeting human motion onto a humanoid with different mass distribution and actuator bandwidth is lossy in ways nobody has quantified, and graphics-quality motion is not the same thing as reliable task completion.
2.7 Jiajun Wu's group — neuro-symbolic and physical scene understanding
Jiajun Wu runs the strongest intellectual counterargument to pure VLA scaling at Stanford. His position: intelligence requires the right level of abstraction — programs, symbols, object intrinsics — not end-to-end pixels.
The programmatic statement is Building Intelligent Agents with Neuro-Symbolic Concepts (with Jiayuan Mao and Josh Tenenbaum, CACM 2026). Around it: The Scene Language (CVPR 2025), representing scenes as programs plus words plus embeddings; What Makes a Maze Look Like a Maze? (ICLR 2025); FluidNexus and Birth and Death of a Rose (CVPR 2025) on 4D and fluid physical modelling; DexSkin (CoRL 2025), a conformable robot skin for contact-rich manipulation; TWIST (with Karen Liu); and a Science Robotics 2025 review of learning-based dynamics models for manipulation.
Data stance: mostly simulation and vision; robot work via collaboration.
Why promising: compositional, program-based representations are interpretable and editable in ways monolithic policies are not, and if VLA scaling stalls this is the most developed alternative.
Skeptic's view: neuro-symbolic approaches have repeatedly lost to scale, and much of the output is graphics-adjacent — it is not always clear how it reaches a robot's control loop.
2.8 Autonomous Systems Lab — Marco Pavone
ASL argues that autonomy is only useful if it is provably safe, and pairs learned components with reachability analysis, conformal prediction and runtime monitoring. Pavone is simultaneously Senior Director of AI Research at NVIDIA, which shapes the lab's output toward autonomous vehicles and simulation.
The manipulation-relevant work is about auditing learned policies rather than building them: CUPID: Curating Data your Robot Loves with Influence Functions (CoRL 2025) identifies which training demonstrations actually help; RoboMonkey (CoRL 2025) scales test-time sampling and verification for VLAs; Real-Time Out-of-Distribution Failure Prevention via Multi-Modal Reasoning (CoRL 2025) detects when a policy has left its competence; Unpacking Failure Modes of Generative Policies does the diagnostic work. With Cutkosky's lab, ASL co-authored ReachBot — Locomotion as manipulation with ReachBot, Science Robotics 2024.
Data stance: simulation and formal analysis over collection. Data is something to be audited, not merely accumulated.
Why promising: as VLAs enter cars and spacecraft, runtime out-of-distribution detection and test-time verification become the binding constraint. This is the strongest group anywhere on that.
Skeptic's view: conformal and reachability guarantees rest on exchangeability and model assumptions that real deployments violate. The safety layer risks becoming a conservative wrapper that caps capability rather than a genuine correctness proof.
2.9 Multi-Robot Systems Lab — Mac Schwager
MSL bets that if you give robots a differentiable, photorealistic, uncertainty-aware map — specifically 3D Gaussian Splatting — then mapping, planning, simulation and semantics collapse into one representation.
Splat-Nav (IEEE T-RO 2025) does safe real-time navigation directly inside a Gaussian Splatting map; a companion paper adds a control barrier function for online splat maps (ICRA 2025, with Kennedy). HAMMER extends this to heterogeneous multi-robot semantic splatting. GRaD-Nav and GRaD-Nav++ (2025) fly drones through Gaussian radiance fields with differentiable dynamics. VISTA (RA-L 2026) does open-vocabulary task-relevant exploration; SINGER (ICRA 2026) runs an onboard generalist vision-language navigation policy; Phys2Real (ICRA 2026, with Jiajun Wu) fuses VLM priors with online system identification; and π, But Make It Fly (2026) transfers VLA models to aerial manipulation.
Data stance: build the map online from the robot's own sensors, then simulate inside the reconstruction rather than in a hand-authored scene.
Why promising: real-to-sim-to-real via splatting is one of the few genuinely new capabilities of the last two years, and it applies directly to the evaluation problem.
Skeptic's view: Gaussian Splatting is a view-synthesis representation, not a geometric or physical one. Free-space and collision guarantees derived from it inherit every reconstruction artifact, and it degrades badly under motion blur, dynamic scenes and sparse coverage.
2.10 CHARM Lab — Allison Okamura
The Collaborative Haptics and Robotics in Medicine lab works on the claim that touch is underexploited in both directions: robots need it to act safely in contact with people, and people need it to teleoperate robots well.
Research spans haptics for teleoperation and VR; wearable haptic interfaces using skin stretch and vibrotactile feedback; steerable needles for percutaneous intervention; human motor performance in robot-assisted surgery; rehabilitation robotics; and soft growing "vine" robots that move by tip eversion rather than locomotion. Recent output includes a hermetic transparent vine robot for pipe inspection deployed in a live wastewater pipe (RoboSoft 2026), self-fastening anchors for auto-wearable robots, a miniaturised pneumatic actuator array for deep-pressure tactile stimulation (Haptics Symposium 2026), and phloSAR, a portable pneumatic supply that untethers soft robots.
Data stance: human-subjects psychophysics and controlled user studies. The measured quantity is human perception and performance, not policy success rate.
Why promising: teleoperation is how almost all robot manipulation data is currently collected, so improvements to the human interface improve everyone else's dataset. Vine robots are also genuinely cheap — polyethylene tubing — and reach places no rigid robot can.
Skeptic's view: the field's flagship claim, that force feedback reliably improves surgical outcomes, has been contested for two decades, and wearable haptic devices have repeatedly failed to leave the lab.
2.11 Biomimetics and Dexterous Manipulation Lab — Mark Cutkosky
BDML copies the mechanism, not the appearance: controlled dry adhesion, tuned compliance and passive dynamics let simple hardware achieve what complex control cannot.
The most consequential line is gecko-inspired dry adhesives — Gentle Grasping With Gecko-Inspired Adhesives in Extreme Environments (IEEE Transactions on Field Robotics 2025) and a gecko-adhesive lasso for de-tumbling orbital debris (iSpaRo 2025), building on the Astrobee gripper flown on the ISS. ReachBot, the extending-boom limbed robot for Martian lava tubes, appeared in Science Robotics 2024 under the framing "locomotion as manipulation." Other lines: long-reach robotic manipulation for lunar assembly (iSpaRo 2025 best paper), tactile-informed action primitives for clutter (ICRA 2024, with Bohg), multi-modal jumping/flying/perching robots, and MRI-compatible force-sensing surgical tools.
Data stance: essentially none. This is a mechanism-design and physical-modelling lab; learning enters only through collaborations.
Why promising: adhesion and long-reach manipulation solve space and extreme-environment problems that no amount of data solves, and space robotics is a real, funded market.
Skeptic's view: gecko adhesives have been "two years from deployment" since roughly 2008, and performance collapses on dusty, rough or regolith-covered surfaces — precisely the lunar and Martian case.
2.12 Stanford Robotics Lab — Oussama Khatib
Khatib's lab is the intellectual origin of much of what humanoid companies run underneath their learned policies. The operational space formulation — unified task and posture control with contact and force at the centre — dates to his early Stanford work and remains the standard substrate for whole-body control.
The flagship system is OceanOneK, a bimanual underwater humanoid with stereo vision, underactuated four-fingered hands and bilateral haptic feedback, rated to 1,000 m. Its 2022 Mediterranean campaign reached the submarine Le Protée at 124 m, a Roman wreck near Aléria at 334 m, the Francesco Crispi at 507 m, and a record 852 m seafloor touch — the deepest a humanoid has reached. Other lines: multi-contact force control, torque-transformer control of position-controlled arms, and the SAI 2.0 / OpenSai open-source simulation and control stack now being pushed through the Robotics Center.
Data stance: model-based and essentially anti-data. Dynamics and control theory, with the human supplying intent via haptic teleoperation.
Why promising: haptic telepresence for deep-sea archaeology is a genuinely deployed capability, and operational-space control is load-bearing infrastructure for the entire humanoid industry.
Skeptic's view: the lab's peer-reviewed output has thinned markedly since about 2022, and OceanOneK is teleoperated — impressive engineering that sidesteps the autonomy question the rest of the field is fighting over.
2.13 ARM Lab — Monroe Kennedy III
The Assistive Robotics and Manipulation Lab argues that a robot working alongside or on behalf of a person needs three things: high-resolution touch, an explicit model of the human partner, and interfaces usable by non-experts.
Its signature hardware is the DenseTact optical tactile sensor family, extended by TensorTouch (IEEE T-RO, April 2026), which calibrates tactile sensors to recover full high-resolution stress tensors, and DOT-Sim (ICRA 2026, with Rika Antonova and Leo Guibas), a differentiable optical tactile simulator for real-to-sim calibration. Around it: Next Best Sense (ICRA 2025) guiding 3D Gaussian Splatting with touch; Splat-MOVER (CoRL 2024); DexFruit (RA-L 2025) for gentle fruit handling; ProACT (IEEE TNSRE 2025), an AR testbed for intelligent prosthetic arms; Towards Accessible Robot Control (2026) comparing kinesthetic and mixed-reality teleoperation; and Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models (2026, with Schwager) — an early attempt to apply LLM interpretability tooling to robot policies.
Data stance: real-robot multimodal data with a tight sim-calibration loop. Build the sensor, then make the simulator match it.
Why promising: touch is the most obvious missing modality in current VLAs, and Kennedy is one of very few people building both the sensor and its differentiable simulator.
Skeptic's view: optical tactile sensors remain fragile, low-bandwidth and hard to manufacture consistently, and nobody has yet shown that adding touch produces a step change in generalist policy performance rather than a task-specific gain.
2.14 Biomechatronics and Neuromuscular Biomechanics — Steve Collins and Scott Delp
Two Stanford labs supply the human-side physics that humanoid robotics increasingly borrows.
Steve Collins' Biomechatronics Lab holds that you cannot design assistive devices analytically because you cannot model the human — so optimise the human–robot system empirically, per person, in the loop. The methodological capstone is On human-in-the-loop optimization of human–robot interaction (Slade et al., Nature 633:779–788, 2024). The most robotics-transferable artifact is Elastic energy-recycling actuators for efficient robots (Krimsky & Collins, Science Robotics 2024) — because energy, not policy, is the binding constraint on legged robots. Also: a twisted-string ankle exoskeleton for running (2025), and improved CMA-ES for noisy robot optimisation (2026).
Scott Delp's Neuromuscular Biomechanics Lab maintains OpenSim, the field-standard open musculoskeletal simulation stack, and through the Mobilize Center produces large mobility datasets. Recent: the AddBiomechanics dataset of physics-consistent human motion; markerless motion capture enhancement (IEEE TBME 2025); and GaitDynamics, a generative foundation model for walking and running (Nature Biomedical Engineering 2026).
Why they matter here: OpenSim skeletons are the reference for musculoskeletal RL and for retargeting human motion onto humanoids, and AddBiomechanics/GaitDynamics supply physically consistent human motion priors — exactly what whole-body humanoid controllers need.
Skeptic's view: neither lab targets robots. Muscle-level models are far too slow and over-parameterised for real-time humanoid control, and human-in-the-loop optimisation takes hours per subject, which has kept exoskeletons commercially marginal for a decade.
2.15 Soft robotics and mechanisms — Zhao Lab and SHAPE Lab
Renee Zhao's lab builds magnetically actuated origami millirobots. The headline result is the milli-spinner thrombectomy device (Nature 642:336–342, June 2025) — a spinning millirobot that compacts a blood clot rather than grabbing it, reported at substantially higher recanalisation rates than aspiration. Follow-ons include a magnetic milli-spinner for robotic endovascular surgery (Advanced Materials 2026), RodOri elastic-rod origami (Science Advances 2026), and an amphibious electromagnetic soft robot (2025).
Sean Follmer's SHAPE Lab works on shape-changing interfaces, haptics and accessibility — the human-interface end of the same problem CHARM attacks.
Why they matter here: they are the clearest reminder that "robotics" is not coextensive with "robot learning." A device that dissolves a clot at scale will save more lives than any household manipulation policy of this decade.
Skeptic's view: these are materials and mechanism results with essentially no learning content. They scale poorly into the VLA-dominated agenda, and clinical translation is years away.
2.16 Stanford's industry gravity: HAI, the Digital Economy Lab, and the spin-out pipeline
Stanford HAI runs the field's most useful policy and measurement layer: a robotics topic hub, the annual AI Index with a robotics chapter, and work like QuantiPhy (January 2026), a benchmark showing frontier models fail at basic physical estimates of size, speed and distance — a real embodied-foundation-model contribution.
Erik Brynjolfsson's Digital Economy Lab supplies the only causal numbers on physical automation coming out of Stanford. Minimum Wages and Rise of the Robots (February 2026) uses plant-level robot import data and Census records from 1992–2021 with state-border discontinuities, finding a 10% minimum-wage increase raises robot adoption roughly 8% relative to the mean. The caveat is important: this is about industrial arms, not general-purpose manipulators.
The spin-out pipeline is the dominant fact about Stanford robotics in 2026. Physical Intelligence (Finn, with Hausman and Levine). Sunday Robotics (Tony Zhao of ALOHA, with Cheng Chi of UMI and Diffusion Policy). World Labs (Fei-Fei Li). Deep ties into Google DeepMind (Sadigh, Open X-Embodiment, Gen2Act, RT-Affordance) and NVIDIA (Pavone; Cosmos Policy at ICLR 2026). Industry affiliates at the Robotics Center include Google, Autodesk, Intuitive Surgical and Foxconn.
Skeptic's view: the ties are now tight enough that Stanford increasingly validates industry agendas rather than setting independent ones, and there has been no marquee robot-learning faculty hire since Song in 2023. The best student hardware becomes a startup within eighteen months, and the open version stops being maintained.
2.17 Deep dive: what Stanford actually contributed, 2023–2026
Strip out the press releases and Stanford's contribution to this period is unusually legible. It is three interfaces and one benchmark.
ALOHA proved that a bimanual teleoperation rig costing less than a used car, paired with action chunking, could learn contact-rich bimanual tasks from fifty demonstrations. Before ALOHA, credible manipulation research required a six-figure hardware budget. After it, a dozen labs and half a dozen startups were running the same recipe within a year. Mobile ALOHA extended this to whole-body mobile manipulation and produced the field's first genuinely viral robot video.
UMI went further by removing the robot from the data-collection loop entirely. A 3D-printed handheld gripper and a GoPro produce demonstrations at 111 per hour versus 35 for SpaceMouse teleoperation — about 48% of bare-hand speed — that transfer zero-shot to UR5e and Franka arms. DexUMI extended the same logic to multi-fingered hands with a wearable exoskeleton and video inpainting to close the visual gap. The conceptual move is important and underrated: rather than making robots easier to drive, make human demonstrations already be in the robot's action space.
TWIST2 applied the same trick to whole-body humanoid data — mocap-free, VR-based, a $250 robot neck, a claimed hundred demonstrations in fifteen minutes. If humanoids matter, whole-body data collection is the next bottleneck and this is the cheapest known answer.
BEHAVIOR is the benchmark. Its bet is that the field's evaluation crisis (Chapter 13) is best solved by making the tasks long, realistic and partially creditable, in a simulator good enough to model cloth and fluids. Whether simulation rankings predict real rankings remains open — but BEHAVIOR is the only effort seriously trying to find out at household scale.
What Stanford did not produce in this period is a frontier model. π0 came out of Physical Intelligence, Gemini Robotics out of DeepMind, GR00T out of NVIDIA. The university's comparative advantage turned out to be the layer beneath the models — the hardware, the interfaces, the benchmarks, and the people. Whether that is a healthy division of labour or a slow hollowing-out is the argument running through the rest of this book.