The board

14 companies 4 axes 30 placements italics = what the bet costs tap a name for the source

01 Scale the dataWhere competence comes from 12

its own trials

the most expensive route. It needs real robots practising, on real time

Physical Intelligence

Physical IntelligenceDemonstrations, then expert corrections, then RL from its own autonomous attempts. from Levine, Berkeley + Finn, Stanford source

a human at the controls

an hour of human time per hour of data

Physical IntelligenceToyota Research InstituteGoogle DeepMind

Physical IntelligenceOver 10,000 hours of its own teleoperated demonstrations, seven robot configurations, 68 tasks. from Levine, Berkeley + Finn, Stanford source

Toyota Research Institute468 hours of its own bimanual teleop plus 1,150 hours of public robot data. Cameras and proprioception only, no force. from Tedrake, MIT source

Google DeepMindMulti-embodiment robot data collected on ALOHA-2, bi-arm Franka and the Apollo humanoid. source

a glove, and no robot

the arm behind the hand is still a human arm

Sunday Robotics

Sunday RoboticsA Skill Capture Glove with the exact same geometry and sensor layout as the robot hand, to remove the translation gap. from Chi and Zhao, from Song's group at Columbia and Stanford source

humans on video

no true actions, ever. Video shows what happened, never what was commanded

NVIDIADyna RoboticsSkild AIRhoda AI

NVIDIAGR00T N1.7 pretrains on EgoScale: 20,854 hours of human egocentric video across twenty-plus task categories. source

Dyna RoboticsA world-action model pre-trained on over a million hours of egocentric human video, actions inferred from 3D hand-pose tracks. source

Skild AIPre-trains on internet human video, on the view that humans are another form factor. from Pathak and Gupta, CMU source

Rhoda AIPretrains on hundreds of millions of web videos, then asks for one to ten hours of trajectory data per new task. source

a simulator

wrong in exactly the place contact happens: friction, deformation, compliance

Eka RoboticsGenesis AISkild AI

Eka RoboticsHigh-fidelity simulation and tactile grippers, explicitly instead of human imitation data. from Agrawal, MIT + Haarnoja, ex-DeepMind source

Genesis AIAn in-house simulation stack generating synthetic data at scale, paired with real-robot collection. A dual engine. from Zhou Xian, CEO; the Genesis engine came from a 20-plus-lab collaboration source

Skild AINames large-scale simulation alongside internet video as the two pre-training sources, simulation first. from Pathak and Gupta, CMU source

02 Scale the representationWhat the model actually predicts 7

language as the bridge

language underdetermines force. You cannot say how hard to squeeze

Google DeepMindPhysical Intelligence

Google DeepMindGemini Robotics 2 pairs a thinking VLA with ER 2, an embodied-reasoning model that plans before it acts. source

Physical IntelligenceA vision-language-action model emitting continuous action chunks by flow matching. from Levine, Berkeley + Finn, Stanford source

force as the native channel

new hardware, and nothing pretrained to start from

Eka Robotics

Eka RoboticsForce, not language, as the native representation of physical interaction. from Agrawal, MIT + Haarnoja, ex-DeepMind source

future video and action, jointly

predicting pixels well is not the same as controlling well

Dyna RoboticsRhoda AI

Dyna RoboticsDyna-2 predicts future frames and robot actions together, as one world-action model. source

Rhoda AIA Direct Video Action model: forecast the video, act on the forecast. Measured on one real job, and they say so. source

latent actions, inferred not labelled

the latent space is unlabelled, so nobody can tell you what it learned

AgiBot

AgiBotGO-2 is a ViLLA model, vision-language-latent-action, bridging planning and execution. source

physics and uncertainty, not language

a belief about risk is harder to demo than a robot following an instruction

Field AI

Field AIPhysics-first foundation models, intrinsically risk-aware, navigating with no prior maps, GPS or predefined paths. source

03 Scale the tasksHow a new task arrives 6

many tasks, one model

interference. A new task quietly degrades the old ones

Physical IntelligenceSkild AIEka Robotics

Physical IntelligenceOne policy across many jobs rather than one policy per job. from Levine, Berkeley + Finn, Stanford source

Skild AIGenerality across task and hardware is the stated product, not a side effect. from Pathak and Gupta, CMU source

Eka RoboticsSays robotics faced a tradeoff, generality or speed, and that the real world requires both. from Agrawal, MIT + Haarnoja, ex-DeepMind source

one demonstration, no retraining

in-context skills stay more brittle than finetuned ones, by their own admission

Generalist AI

Generalist AI59% one-shot across ten tasks with no gradient updates. 83% after ten gradient steps on five minutes of new data. source

plan it, then execute it

a planner that is wrong is confidently wrong for a long time

AgiBot

AgiBotGO-2 bridges planning and execution with an action chain-of-thought, for long-horizon work. source

reliability as the headline number

says nothing about the second task, or the fortieth

Dyna Robotics

Dyna RoboticsDYNA-1 folded over 800 napkins across 24 hours at 99.4%, zero interventions, 60% of human throughput. source

04 Scale the bodiesWhat the model is allowed to run on 5

one policy, many bodies

nothing is tuned for anything, and the last few percent is where deployment lives

NVIDIASkild AIField AIGoogle DeepMind

NVIDIAGR00T N1 shipped as the first open humanoid foundation model, then opened to every developer. source

Skild AIOne omni-bodied model for quadrupeds, humanoids, tabletop arms and mobile manipulators. from Pathak and Gupta, CMU source

Field AIOne autonomy across humanoids, quadrupeds, wheeled robots and passenger-scale vehicles. source

Google DeepMindGemini Robotics 2 adapts to a new bi-arm embodiment in hours, typically under 200 examples. The SO-101 is one of them. source

one body, mass-manufactured

you must be right about the design before you tool up

Figure AI

Figure AIOne Figure 03 per hour at BotQ, up from one a day in under 120 days. A 24x ramp. source

Every company, and the axes it is betting on