Comparison Tables

A.1 Academic labs at a glance

Lab	Institution	PI	Core bet	Data stance
SVL / BEHAVIOR	Stanford	Fei-Fei Li, Jiajun Wu	Benchmark-first household AI	Sim-first, teleop-heavy
IPRL	Stanford	Jeannette Bohg	Interactive perception; force is primary	Pluralist; edits data across sources
ILIAD	Stanford	Dorsa Sadigh	Humans in the loop; curation	Real teleop, curated
IRIS	Stanford	Chelsea Finn	Cheap hardware + imitation scaling	Real-world first
REAL	Stanford	Shuran Song	Embodiment-agnostic collection interfaces	In-the-wild, open hardware
Movement Lab	Stanford	C. Karen Liu	Humanoids = physics-based characters	Human mocap → retarget → sim
ASL	Stanford	Marco Pavone	Provable safety around learned parts	Audit data, don't just collect it
MSL	Stanford	Mac Schwager	Gaussian splatting as universal map	Build the map from the robot's own sensors
CHARM	Stanford	Allison Okamura	Touch in both directions	Human-subjects psychophysics
BDML	Stanford	Mark Cutkosky	Mechanism over algorithm	None (mechanism lab)
Stanford Robotics Lab	Stanford	Oussama Khatib	Operational-space control; telepresence	Model-based, anti-data
ARM Lab	Stanford	Monroe Kennedy III	Tactile + collaboration	Real multimodal + sim calibration
Robot Locomotion	MIT	Russ Tedrake	Rigour + behaviour cloning at fleet scale	Teleop-first
Improbable AI	MIT	Pulkit Agrawal	Dexterity is a force problem	Sim + perioperation
LIS	MIT	Kaelbling, Lozano-Pérez	Abstraction beats scale	Explicitly anti-scaling
Distributed Robotics	MIT	Daniela Rus	Morphology + compact networks	Pro-data
GelSight	MIT	Ted Adelson	Touch as vision	Hardware-first
Biomimetic Robotics	MIT	Sangbae Kim (leave)	Proprioceptive actuators	Model-based
Interactive Robotics	MIT	Julie Shah	Robots as teammates	Low-N human-in-the-loop
SPARK	MIT	Luca Carlone	Certifiable perception, scene graphs	Hybrid
Robust Robotics	MIT	Nicholas Roy	Autonomy under uncertainty	Foundation models as evidence, not truth
CDFG	MIT	Wojciech Matusik	Co-design body and controller	Differentiable sim + fabrication
d'Arbeloff	MIT	Harry Asada	Wearable and supernumerary robots	Model-based
Biomechatronics	MIT	Hugh Herr	Redesign the body for the machine	Clinical, n-of-few

A.2 Companies by thesis

Thesis	Companies
Vertically integrated humanoid	Figure, Tesla, 1X, XPeng, Apptronik (hardware side)
Hardware-agnostic brain	Physical Intelligence, Skild, Generalist, Google DeepMind, Field AI, Sanctuary (post-pivot)
Wheeled/semi-humanoid pragmatism	Walden, Galbot, Dexterity, Cobot, Agility, Ai² Robotics
Cheap hardware, outsourced intelligence	Unitree, Fourier (partly), Kepler, Astribot
Deployment flywheel	Dexterity, Ambi, Chef, Dyna, RightHand, Path, Amazon
Simulation-first	NVIDIA, Skild, Galbot, Genesis AI
Human-data-first	Generalist, Sunday, Meta, Skild
Platform / picks and shovels	NVIDIA, Hugging Face, Scale AI, Encord, LG/Samsung/Hyundai Mobis (actuators)

A.3 The data collection cost/fidelity frontier

	Cheap	Expensive
Low fidelity	Internet video, Ego4D	—
Medium fidelity	Simulation, Aria/EgoDex wearables	Motion capture studios
High fidelity	UMI / DexUMI / DEXOP / Skill Capture	Teleoperation fleets, AgiBot data factory

The bottom-left cell is where the interesting engineering is. Everything in Chapter 9 is an attempt to move data from the top-left or bottom-right into the bottom-left.

---