Nine Competing Theses, Taxonomised

11.1 End-to-end VLA scaling

Claim: one network from pixels and language to actions; scale data and diversity, and structure emerges. Adherents: Physical Intelligence, Google DeepMind, the Octo/OpenVLA community, Generalist AI. Best evidence: π0.7 showing "the first signs of compositional generalization," including folding laundry on a robot with no laundry-folding data; Open X-Embodiment positive transfer; the imitation scaling laws paper. Best critique: Brooks' "end-to-end is a myth" argument — every great scaling success had an engineered front end. And π0.7's gains came substantially from prompt structure engineering (textual descriptions, subgoal images, speed and quality metadata), which is hand-designed inductive bias wearing a scaling costume.

11.2 Hierarchical VLA (System 1 / System 2)

Claim: semantics and control need different clock rates, so split them. Adherents: Figure, NVIDIA, Google DeepMind, most humanoid companies in practice. Best evidence: Helix — a 7B VLM at 7–9 Hz over an 80M policy at 200 Hz, "because VLM backbones are general, but not fast, and robot visuomotor policies are fast but not general." Helix 02 added S0, a 10M network at 1 kHz replacing 109,504 lines of hand-written C++, producing a four-minute, 61-action autonomous dishwasher sequence. GR00T N1 uses the same split. Best critique: the interface between layers is a hand-designed bottleneck, latent handoffs are hard to debug, and the split may be a temporary artifact of inference cost rather than a principle. If edge inference gets 100× cheaper, the argument for it largely evaporates.

11.3 World-model / video-prediction-first

Claim: learn a predictive simulator from video, then plan, evaluate or generate data inside it. Adherents: Meta (V-JEPA 2), Google DeepMind (Genie 3), 1X World Model Lab, World Labs, X Square Robot, NVIDIA Cosmos. Best evidence: V-JEPA 2's 62-hours-of-robot-data result; Genie 3 at 720p/24fps with minute-scale consistency, explicitly pitched at training agents; and — the most practically interesting framing — 1X's reframing of the world model as an evaluation substrate: "if you train a robot to perform 1000 unique tasks, it is very hard to know whether a new model has made the robot better at all 1000." Best critique: Levine's "Language Models in Plato's Cave" — video models had strictly more information than LLMs and still lost at reasoning. Predicting pixels is not the same as predicting the consequences of your own actions.

11.4 Classical model-based control with learned components

Claim: keep MPC and whole-body control; learn only the parts that resist modelling. Adherents: Boston Dynamics, Khatib, Sangbae Kim, RAI Institute, Goldberg's "Good Old Fashioned Engineering." Best evidence: Atlas combining RL, reference trajectories and proprioceptive feedback, with a fridge-carrying policy generalising from 50–70 lb training loads to 100+ lb. Goldberg's framing: "modularity, metrics, and step-by-step algorithms based on geometry and physics that can be fully understood and often guaranteed to perform reliably." Best critique: it does not obviously extend to open-world semantics, and Figure's deletion of 109,504 lines of C++ is a datapoint on the other side.

11.5 TAMP and neuro-symbolic

Claim: long-horizon manipulation is a hybrid discrete-continuous search problem that no monolithic policy has the structure to solve. Adherents: Kaelbling, Lozano-Pérez, Garrett, Tom Silver, Jiajun Wu. Best evidence: the 2026 turn in which VLMs supply the predicates — From Pixels to Predicates and Open-World TAMP via VLM-Generated Constraints — solving the symbol-grounding problem that stalled the programme for thirty years. Best critique: demonstrations remain close to tabletop domains, invented predicates inherit VLM hallucination, and Gemini Robotics' "thinking before acting" is arguably TAMP re-implemented in natural language by a competitor with a thousand times the compute.

11.6 Tactile-first / force-centric

Claim: vision cannot supply the signal dexterity needs; touch is not an add-on. Adherents: Brooks (as critic), Adelson, Pulkit Agrawal, Monroe Kennedy, Meta FAIR, Amazon (Vulcan), Figure (partially). Best evidence: Meta's Digit 360 (8M+ taxels, 1 mN sensitivity) and Sparsh (460k tactile images) as the first general-purpose tactile encoder; DEXOP capturing full-hand tactile at collection time; Amazon's Vulcan handling ~75% of stowed item types via force feedback; Figure 03's fingertip sensors detecting 3 grams. Best critique: no touch corpus exists at any meaningful scale, there is no standard representation or interchange format, sensors remain fragile, and nobody has yet demonstrated that a tactile-conditioned policy beats a vision-only one on the same task with the same data budget.

11.7 Morphology-first

Claim: better hands, wrists and actuators unlock more than better models. Adherents: 1X (25-DOF tendon hands), Apptronik (22-DOF SharpaWave), Sangbae Kim, Cutkosky, Asada, Herr, Boston Dynamics. Best evidence: proprioceptive actuators made dynamic legged locomotion tractable; Asada's loop-closure grasping makes a grip strong and gentle simultaneously; Herr's AMI restores proprioception no controller could synthesise. Best critique: Tedrake's, and it is devastating in its simplicity — "the real question is just durability... I have not seen a more dexterous hand that could have done the work our hand has done." Degrees of freedom without data are inert, and complex hands are reliably the least reliable subsystem.

11.8 Teleoperation-as-product

Claim: ship now with humans in the loop; supervision cost falls as autonomy improves. Adherents: 1X ("Scheduled Expert Mode"), Nimble, Plus One Robotics, most warehouse exception-handling. Best evidence: Plus One's CTO Shaun Edwards notes "one person can supervise 50 or more robots at once"; and every intervention is a labelled demonstration, so the supervision cost buys training data. Best critique: the privacy problem is unsolved for homes — 1X's own FAQ has to answer whether an operator can enter your NEO at will — and the margins invert only if autonomy actually arrives on schedule, which is the assumption under test.

11.9 Simulation-first

Claim: generate the data you cannot collect. Adherents: NVIDIA, Skild, Galbot, Genesis AI, Boston Dynamics (for locomotion), Isola (for appearance). Best evidence: Skild's 100,000-body universe; Helix 02's zero-shot sim-to-real stair traversal; LucidSim's zero-real-image parkour transfer; DrEureka's automated randomisation. Best critique: Goldberg on contact, and Levine on discrepancy-fitting. Simulation solves appearance and kinematics; it does not solve friction, deformation or compliance, and those are what manipulation is made of.

11.10 A note on how to read a disagreement

A useful diagnostic when reading any claim in this field: ask which of the nine theses the speaker's revenue depends on. NVIDIA benefits if simulation-first is true. Meta benefits if egocentric video is sufficient. Figure benefits if vertical integration is necessary. Unitree benefits if the brain is a commodity. Tedrake benefits if wheels are enough.

This is not an accusation of bad faith — everyone in this book is intellectually serious, and several of them argue against their own commercial interest (Levine on real data, Tedrake on legs, Goldberg on data scarcity while running a company that benefits from scarcity being solvable). But the correlation between thesis and business model is high enough that it should be the second thing you check, after the evidence.

---