Competing Theses

Modern robotics has not converged. Underneath the shared vocabulary of foundation models and generalist policies, the field is running at least six different bets about what gets built next, and the bets are expensive enough that labs must choose: NVIDIA pretrained Cosmos 3 on up to 31 trillion tokens, using clusters as large as 2,048 GB200 GPUs NVIDIA 2026↓Cosmos 3: Omnimodal World Models for Physical AINVIDIA, Aditi, Niket Agarwal et al., 2026, and Physical Intelligence has staked its whole program, from π0 to π0.7, on the scaling thesis Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. What keeps these from being taste debates is that each thesis, stated precisely, predicts something the world can check. This module states the six theses, names who is betting on each, gives the strongest published evidence on both sides, and records the observation that would falsify each one.

The six theses

The table below is the map. Each row is one thesis with its core claim, its named proponents, and a compact version of its falsification signal. Switch the selection from end-to-end VLA scaling to world-model-based training to read the evidence for and against, with citations, and the full falsification criterion.

Two cautions before the details. First, the theses are not mutually exclusive, and the current frontier systems are already hybrids: π0.7 is an end-to-end vision-language-action modelvision-language-action modelA robot policy built by adapting a vision-language model, pretrained on web-scale image and text data, to output robot actions. RT-2 established the approach by representing actions as text tokens and co-fine-tuning on web and robot data together, showing that semantic knowledge from web pretraining transfers into control. The same recipe underlies OpenVLA and the π series of generalist policies. that also reasons hierarchically, can receive optional subgoal-image prompts from a separate BAGEL-initialized world modelworld modelA model of how an environment evolves that an agent can query to make decisions: given the current observation or state and a candidate action, it predicts what happens next, so a policy can be trained, evaluated, or planned against the model instead of the real world. The 2026 robotics survey draws the functional line: producing plausible future images is not enough, because a system qualifies only if its predictions change under the agent's action in a way that is useful for decision-making. The single name covers at least six architecturally distinct paradigms, from compact latent dynamics models to action-conditioned video generators., and accepts RL post-training Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. The disagreement is about where the marginal bet belongs, not about which single architecture wins outright. Second, every falsification criterion below depends on measurement the field cannot yet do well; The Evaluation Crisis is the reason a claim like "10x more data, no gain" is harder to establish than it sounds.

End-to-end VLA scaling

The claim: keep the architecture general, scale data and parameters, and generalist robot intelligence falls out the way generalist language modeling did. The intellectual frame is Rich Sutton's 2019 essay The Bitter Lesson, the argument that general methods leveraging computation beat hand-built structure over the long run Sutton 2019↓The Bitter LessonRich Sutton, 2019; Brooks answered six days later, in a reply that paraphrases Sutton and describes itself, including its closing comment, as "seventy six words shorter than Sutton’s post" Brooks 2019↓A Better LessonRodney Brooks, 2019. Physical Intelligence's π line is the clearest published arc, NVIDIA's EgoScale team supplied the cleanest quantitative evidence, and Bessemer's investor analysis stakes the capital side: scaling laws are emerging, and data is the moat Levine 2026↓Bessemer Predicts: Robotics and physical AIJeremy Levine, Talia Goldberg, Janelle Teng Wade et al., 2026.

The evidence for is real and recent. EgoScale fit a log-linear scaling lawscaling lawAn empirical regularity between a system's performance and the resources it consumes, parameters, data, or compute, first made precise for language models and now measured for robot learning. Lin and colleagues fit imitation-learning success to the training data and found that generalization to new objects and environments tracks data diversity rather than raw hours. EgoScale extended the measurement to egocentric human video, reporting a log-linear relationship between hours of human data and dexterous-manipulation success across four orders of magnitude., R² = 0.9983, from egocentric human-video pretraining hours to downstream robot performance, and the full recipe lifts average success by 54% on a 22-DoF hand Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. π0.5 cleaned kitchens in homes it had never seen, and π0.7 recombined trained skills into untrained tasks, exactly the emergent behavior the scaling camp predicts Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025 Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. The evidence against is structural. Brooks answered Sutton directly in his 2019 essay: he argues that image-labeling systems use a human-designed convolutional front end for translational invariance Brooks 2019↓A Better LessonRodney Brooks, 2019. His 2025 essay extends that objection to speech preprocessing with roots in telephone engineering, listing FFTs, Mel filter banks and other transformations while emphasizing that implementations use different selections. He also acknowledges that leaving as much as possible to learning was critical to the gains in speech and image tasks Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. On compute cost he is blunter, and in his own words: "we can not afford to put even the results of machine learning (let alone the actual learning) on many of our small robots–self driving cars require about 2,500 Watts of power for computation–a human brain only requires 20 Watts" Brooks 2019↓A Better LessonRodney Brooks, 2019. His 2025 essay adds the tactile argument, covered in Dexterity: vision-only training gives a robot no force channel, and no amount of video supplies one Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. Ken Goldberg quantifies the data problem as a 100,000-year gap and argues good old-fashioned engineering can bootstrap robots into the data collection that closes it Goldberg 2025↓Good old-fashioned engineering can close the 100,000-year "data gap" in roboticsKen Goldberg, Science Robotics, 2025, and Karcini and co-authors argue policy scaling alone misses the interfaces that turn physical data into supervision at all Karcini 2026↓Robots Need More than VLA and World ModelsElis Karcini, Faisal Mehrban, Quang Nguyen et al., 2026. The kill condition is the cleanest of the six: a plateau in real-world success rate despite a 10x data increase with no architectural change. The one caveat, documented in Generalization, is that the measured law so far governs a proxy loss, not success rate across unseen environments.

A hierarchical planner over skills

The bet here is division of labor: a vision-language model reasons about goals and decomposes them into subtasks, and learned skills or a low-level policy execute the plan. As Hierarchical Approaches documents, the 2026 version of this thesis is not the 2022 SayCan architecture, where a separate LLM dispatched a fixed skill library Ahn 2022↓Do As I Can, Not As I Say: Grounding Language in Robotic AffordancesMichael Ahn, Anthony Brohan, Noah Brown et al., 2022; the hierarchy has moved inside the network or into a paired fast system. The named proponents are Google DeepMind, whose Gemini Robotics ER 2 is explicitly built as the high-level brain that plans multi-step tasks and hands motor execution to a lower-level VLA Google DeepMind 2026↓Gemini Robotics ER 2Google DeepMind, 2026; Figure, which describes Helix 02 as a hierarchy in which S2 interprets scenes and language and sequences behaviors, S1 converts perception into full-body joint targets at 200 Hz, and S0 executes at 1 kHz as the whole-body controllerwhole-body controlTreating all of a robot's actuated degrees of freedom, legs, torso, arms, and hands, as one coupled control problem instead of stacking an arm controller on top of a separate locomotion controller. The model-based version optimizes against the full rigid-body dynamics at once: Zhang and colleagues ran whole-body model-predictive control in real time on hardware with plain iLQR and MuJoCo dynamics, across dynamic quadruped locomotion and full-sized humanoid walking. The learned version reframes the same problem as tracking a retargeted human motion, the route H2O took for real-time whole-body teleoperation, with reinforcement learning supplying the balance and contact feasibility that raw retargeting loses. for balance, contact, and coordination Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026; and NVIDIA, whose GR00T line keeps the same dual-system shape.

The strongest evidence for the thesis is long-horizon behavior. Figure reports that Helix 02 completed a continuous four-minute dishwasher unloading-and-reloading task without resets or human intervention, sequencing 61 loco-manipulation actions that it describes as ordered correctly with implicit error recovery Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026, and π0.7 absorbs new tasks through the top of its hierarchy: language coaching fine-tunes the high-level policy, which then runs the task autonomously Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. The evidence against comes from the same family of results. π0.7's compositional generalization emerges from one network, suggesting the skill library can be implicit rather than engineered Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026, and the original separate-planner lineage already demonstrated the brittleness of the handoff: a planner that can only choose among trained skills stalls the moment the scene calls for none of them Ahn 2022↓Do As I Can, Not As I Say: Grounding Language in Robotic AffordancesMichael Ahn, Anthony Brohan, Noah Brown et al., 2022. The thesis dies if a flat end-to-end model, with no high-level inference pass at all, outperforms the best hierarchical system on long-horizon tasks.

World-model-based training

The claim is that a learned model of how the world evolves can replace most real-robot data: plan in it, or train the policy inside it, and the data bottleneck stops binding. NVIDIA's Cosmos 3 and Meta's V-JEPA 2 are the largest such bets; the longer Dreamer lineage is covered in What Is a World Model?.

The case for it is sample efficiency. V-JEPA 2 plans zero-shot reaching, grasping, and pick-and-place on Franka arms in labs it never saw, after action-conditioned post-training on less than 62 hours of unlabeled robot video Assran 2025↓V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and PlanningMahmoud Assran, Adrien Bardes, David Fan et al., 2025. And the idea is already load-bearing inside the scaling camp's flagship: π0.7 can receive optional subgoal-image prompts from a separate world model initialized from BAGEL, which the paper describes as lightweight Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. The case against is cost and fidelity. Cosmos 3 was pretrained on up to 31 trillion tokens, on clusters as large as 2,048 GB200 GPUs NVIDIA 2026↓Cosmos 3: Omnimodal World Models for Physical AINVIDIA, Aditi, Niket Agarwal et al., 2026, and no world model has yet demonstrated physics predictions accurate enough for contact-rich manipulation, where a one-millimeter error flips stick to slip. Karcini and co-authors argue world models alone are insufficient for the supervision problem regardless Karcini 2026↓Robots Need More than VLA and World ModelsElis Karcini, Faisal Mehrban, Quang Nguyen et al., 2026. It dies if world-model training fails to beat VLA scaling alone on real-world contact-rich tasks.

RL fine-tuning on imitation

The claim here has the strongest current numbers behind it: pretrain on demonstrations, then close the reliability gap with reinforcement learning on the real robot. π*0.6 with Recap more than doubled throughput and reached the 90%+ success range on all tasks except diverse laundry after RL on the robot's own experience; on laundry the first Recap iteration already lifts the success rate over 90%, and box assembly finishes at about 90% for folding and labeling Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025. RL-100 reports 100% success across 1,000 evaluation episodes on eight tasks, including a juicing robot that served customers for about seven hours without failure Lei 2025↓RL-100: Performant Robotic Manipulation with Real-World Reinforcement LearningKun Lei, Huanyu Li, Dongjie Yu et al., 2025. ENPIRE goes further and hands the improvement loop itself to coding agents, reaching 99% on dexterous tasks such as pin-box organizing Xiao 2026↓ENPIRE: Agentic Robot Policy Self-Improvement in the Real WorldWenli Xiao, Jia Xie, Tonghe Zhang et al., 2026. The mechanics behind these results are covered in RL Fine-Tuning of Policies.

The evidence against is about scope, not efficacy. Every one of these results is per-task: hours of on-robot training buy a specialist, and neither paper claims the improvement transfers to unseen tasks Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025 Lei 2025↓RL-100: Performant Robotic Manipulation with Real-World Reinforcement LearningKun Lei, Huanyu Li, Dongjie Yu et al., 2025. The generalist comparison is not an exclusivity result: Physical Intelligence reports that one π0.7 model matches the RL-trained specialists on the evaluated laundry-folding, espresso-making, and box-building tasks, sometimes exceeding their throughput, after distilling Recap experience with strategy metadata. The paper measures throughput as successful episodes per hour, not inference speed Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026 Ai 2026↓π0.7: a Steerable Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. The recipe stops being the path and becomes a niche tool if RL-finetuned specialists consistently fail to generalize without another round of per-task training.

Teleoperation as a bridge

The most pragmatic of the six: deploy teleoperated or human-supervised robots for real work now, harvest the data the work generates, and fade the human out as autonomy improves. Nucleus deploys humanoids under human supervision rather than full autonomy on exactly this logic Robots 2026↓Nucleus Robotics Emerges From Stealth With Humanoids on the German Factory FloorRocking Robots, 2026, Physical Intelligence's Recap turns human coaching with corrections into more than doubled throughput Amin 2025↓π*0.6: a VLA That Learns From ExperienceAli Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025, and Bessemer’s investor outlook quotes Voxel51’s Brian Moore and Foxglove’s Adrian Macneil on a data flywheel: “better decisions, better model improvements, and better deployments faster than everyone else.” Bessemer lists both companies in its portfolio. This is an attributed investment thesis, not evidence that deployment revenue funds data collection or guarantees autonomy Levine 2026↓Bessemer Predicts: Robotics and physical AIJeremy Levine, Talia Goldberg, Janelle Teng Wade et al., 2026. The two strongest verified deployment records are both supervised programs of this kind: Agility's 65,000+ operating hours across nine facilities and Figure's 1,250+ hours at BMW Spartanburg Noreika 2026↓Humanoid Robots in 2026: What Is Actually DeployedAlius Noreika, 2026. The rigs that feed this pipeline are compared in Teleoperation Rigs.

The evidence against is that the bridge may not reach the far side. TeleoperationteleoperationDirect remote control of a robot by a human operator, used in robot learning mostly to record expert demonstrations. The ALOHA rig made the setup cheap and precise: the operator back-drives a leader arm that is a joint-for-joint twin of the follower the policy will run on, so the recorded motion is already in the robot's own action space. Mobile ALOHA extended the same leader-follower scheme to a mobile base for whole-body tasks. alone is not a sufficient data strategy in Ian Glow’s view. Bessemer quotes the Zeromatter CEO: "you'll never get the scale or diversity you need from teleop alone." He recommends internet data or simulators with reinforcement learning; Bessemer also cautions that world models are expensive and sim-to-real manipulation remains an open research problem Levine 2026↓Bessemer Predicts: Robotics and physical AIJeremy Levine, Talia Goldberg, Janelle Teng Wade et al., 2026. Meanwhile egocentric human video is demonstrating a cheaper source with a measured scaling law and no robot in the loop at all Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. The bridge dies if autonomous policies reach better than 99% success trained without any teleoperation data, because then it has no reason to exist.

Humanoid versus task-specific

The sixth thesis is a pair, because the two sides falsify each other. The humanoid side, backed by Figure, Tesla, 1X, Apptronik, and Unitree, bets that a human body is the right interface to a world built for humans, so one general form factor addresses every task. The task-specific side, argued by Rodney Brooks, who describes his company's warehouse robots and his work at Robust.AI Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025, and embodied by Agility's logistics-focused Digit, bets that purpose-built systems win on cost and reliability for any task you can name in advance.

The humanoid evidence is the strongest it has ever been. Helix 02 runs whole-body loco-manipulation, including a livestreamed eight-hour autonomous shift on a factory floor Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026 Belmonte 2026↓Figure AI's Helix-02 Robots Complete Full 8-Hour Autonomous Shifts as Humanoid Race IntensifiesKyle Belmonte, 2026, and Gemini Robotics 2 controls full humanoids from feet to fingertips with a single model Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026. The task-specific evidence is that the verified deployment records all belong to narrow applications. In his 2025 essay, Brooks calls plug-compatible humanoids replacing people in manual jobs at lower prices and just as well "pure fantasy thinking" within decades. He also argues that articulated fingers lack the robustness, force and lifetime for industrial applications, and that human-like hand designs have not reached real-world deployment. These are his assessments, not an independently established industry census. His scope matters: he reports earlier factory deployments of Baxter and Sawyer and forecasts differently shaped, task-specialized robots still called humanoids over the next fifteen years Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. Both directions of the kill condition are the same measurement: cost per task. If task-specific robots keep winning it in the same applications, the humanoid thesis dies; if a humanoid reaches parity across a broad task range, the task-specific thesis does.

What would settle it

Read back down the falsification column and a watchlist falls out. A 10x or 100x data increase with no architectural change, measured against real-world success rate, answers the scaling question. A flat model beating the best hierarchical system on long-horizon tasks answers the hierarchy question. A world model that improves real contact-rich performance beyond VLA scaling answers the world-model question. RL specialists transferring without retraining, autonomy above 99% trained with no teleoperation data, and humanoid cost-per-task parity answer the rest. None of these observations is cheap, and several are years away; but each is a concrete, checkable event rather than a matter of taste. Until they arrive, the honest position on all six is the one the evidence column already encodes: the proponents are named, the counter-evidence is on the record, and the field has placed real capital on both sides of every row. The Bear Case takes the capital side of that sentence seriously, and its watchlist tracks the status of each of these observations.