A foundation model is a network trained on broad data at scale and adaptable to many downstream tasksSource 10. In robotics three families carry the label (vision-language-action models, predictive world models and multimodal backbones adapted into either), and the test is whether one checkpoint transfers to new tasks, environments or embodiments after bounded adaptation.

Photographs of the robot arms RoboCat controls, each labelled with its action dimensions: Sawyer and Panda arms with parallel grippers seen in training, and a KUKA arm with a three-finger hand held out for fine-tuning.
An adaptation budget is part of a foundation claim: RoboCat learned Sawyer and Panda arms in training and met the KUKA arm with a three-finger hand only at fine-tuning, from 100 to 1,000 examples.
Paper figure: Konstantinos Bousmalis et al. / arXiv. Licence: CC BY 4.0.

Try the check below.

Self-check

Which result is the strongest evidence that a robot model is functioning as a foundation model?
Read the reasoning
  • The checkpoint has a large parameter countScale may increase capacity, but it says nothing by itself about reuse or adaptation. A seven-billion-parameter specialist remains a specialist.
  • The model performs well on many training task labelsTraining breadth is relevant, but success on the training distribution does not demonstrate transfer. The claim needs a downstream setting that was not simply another row in the training set.
  • One checkpoint adapts with limited data to new tasks and embodimentsThis is the operational evidence. A shared pretrained representation is useful if it reduces the data or optimization needed across genuinely different downstream settings.

Judge foundation models by the breadth and cost of downstream reuse; parameter count and branding are weak selectors.

What must be specified

A claim names its pretraining data, objective, interface and adaptation budget. Data spans web text and images, human video, robot trajectories and simulation; Open X-Embodiment pools 22 robots from 21 institutionsSource 11. Objectives include next-token prediction, contrastive learning, video prediction and action denoising.

Budgets run from zero-shot prompting through parameter-efficient and full fine-tuning to new robot data at scale. RoboCat adapts to new tasks and robots from 100 to 1,000 examplesSource 12, Octo fine-tunes to new sensors and action spaces in hours on consumer GPUsSource 13, and π*0.6's RL from experience more than doubles throughput on some of the hardest tasksSource 14. The newest rung is in-context learning: Generalist reports GEN-1.5 learning a task from one 3 to 12 second demonstration with no trainingSource 15.

Robota-Embed, from Yunze Liu, Linxi Fan, Han Cai and colleagues at NVIDIA, MIT, UC Berkeley, Oxford and CMU (October 2026), encodes vision, language and action in one vector, so a robot can retrieve recorded experience as context. Seeding a frozen DreamZero policy with a retrieved video, chosen by a reranking agent, raised its average success on three simulated DROID-Sim tasks from 47.0% to 56.3% over 100 trials per task, while random reference videos lowered it to 36.0%. Without the reranker, its top-1 retrievals reached 50.3% against 43.3% for Omni-Embed-Nemotron-3BSource 32.

VLA foundations

RT-1 argued for task-agnostic training of high-capacity models on diverse robot dataSource 16. RT-2 co-fine-tunes a vision-language model on robot trajectories and internet-scale vision-language tasks, writing actions as reserved tokensSource 1, and OpenVLA made a comparable recipe public on Open X-EmbodimentSource 2. VLAs still struggle with novel robot setups until fine-tuned: OpenVLA-OFT's recipe lifted OpenVLA's average success across four task suites from 76.5% to 97.1%Source 17.

π0 pairs a pretrained VLM with a separate flow-matching expert for continuous action chunksSource 3. π0.5 credits its generalization to co-training on heterogeneous tasksSource 18, and π0.7's authors call their 2026 model a steerable generalist robotic foundation modelSource 4.

Test-time training stores past observations in fast weights that gradient descent updates while the policy runs. RoboTTT (July 2026) extends visuomotor context to 8K timesteps without growing inference latency and averaged a 79% task completion score on long-horizon real-robot tasks, against 42% for the single-step-context GR00T N1.7 baselineSource 9. ECoMEM (Stanford, September 2026) keeps an explicit record of task facts that a learned reader turns into memory tokens for the VLA, reaching 86.1% success on two new real-robot tasks against 8.6% without memorySource 28.

IMPRINT, from Yilin Wu, Homanga Bharadhwaj, Guanya Shi and colleagues at Amazon FAR and Carnegie Mellon (October 2026), trains a latent memory of past camera frames, actions and joint states with the policy, so no rule has to pick which frames to keep. On the nine memory-dependent RMBench simulation tasks, with 100 trials per task, it raised π0.5 from 25.4% to 72.8% average success; EventVLA, the strongest of six memory baselines, reports 67.8%. The recurrent version took 122.8 ms per 50-step action chunk on an L40S GPU, against 91.0 ms for π0.5 without memory and 960 ms for EventVLASource 33.

Humanoid foundations

GR00T N1 combines a vision-language module with a diffusion transformer, trained on real-robot trajectories, human videos and synthetic datasetsSource 5. For a humanoid, morphology, whole-body balance and shared action coordinates set the reuse boundary. Figure pretrained Helix 2.5 from random initialization entirely on Index, its dataset of human behavior; with identical task data, it succeeded on 56% of zero-shot trials in 30 unseen homes, against 9% for a policy trained from scratchSource 29.

Rows of frames from a Fourier GR-1 humanoid robot at a wooden table handing objects between its hands, placing a peach in a bin and moving items onto plates and into baskets.
GR00T N1 runs a vision-language model at 10 Hz over a 120 Hz action module on the GR-1 humanoid. Frames from real rollouts: pretraining evaluations (top) and post-trained pick-and-place and articulated-object tasks (bottom).
Paper figure: NVIDIA et al. / arXiv. Licence: CC BY 4.0.

Predictive foundations

V-JEPA 2 learns representations from over 1M hours of internet video, then trains an action-conditioned stage on fewer than 62 hours of unlabeled DROID videoSource 6. R3M's embedding, pretrained on Ego4D human video, beats from-scratch training by over 20%Source 19.

Two rows of ALOHA robot frames with arrows between them: putting an object on a plate, folding a shirt, putting candies in a bowl and putting candy in a zip bag.
Cosmos Policy fine-tunes a pretrained video model into a robot policy, so a predictive model becomes the controller. Rollouts on four ALOHA tasks, from folding a shirt to putting candy in a zip bag.
Paper figure: Moo Jin Kim et al. / arXiv. Licence: CC BY 4.0.

Cosmos frames a world foundation model as a general-purpose model fine-tuned into custom onesSource 20. Cosmos 3 unifies understanding, generation, forward and inverse dynamics and policy outputs in one omnimodal modelSource 7, and Cosmos Policy turns Cosmos-Predict2 into a policy in one post-training stage, reaching 98.5% on LIBEROSource 21. Each output needs its own test: Ctrl-World ranks policies without real rolloutsSource 22.

DreamZero (February 2026) is a world action model on a pretrained video diffusion backbone that predicts video and actions together; in real-robot tests it more than doubled average task progress on new tasks and environments over state-of-the-art VLAs, and its 14-billion-parameter model ran closed-loop at 7 HzSource 8.

What pretraining does not erase

In TRI's blind, randomized trials, multitask pretraining matched from-scratch policies with 3 to 5 times less dataSource 23. VC-1 found no pretrained visual representation universally dominantSource 24, and Toyota Research Institute's 89-policy study found that robot-only training degrades the backbone's visiolinguistic understandingSource 31. Generalist reports that GEN-1's base model, pretrained on over half a million hours of wearable human data and no robot data, averaged 99% on three tasks with about one hour of robot data per task, against 19% from scratchSource 30.

Pretraining leaves calibration and control integration, closed-loop evaluation under perturbation, latency limits and safety cases to the builder. RoboPAIR jailbroke LLM-controlled robots, including a deployed Unitree Go2, often at 100% attack successSource 25, and a downloadable checkpoint can hide its training mixture, adaptation data or benchmark protocol.

A comparison framework

FamilyPretrained capabilityTypical downstream outputStrongest transfer claimMain unresolved cost
VLAVision-language semanticsAction tokens or continuous chunksNew tasks, objects and instructionsRobot data and control integration
Generalist humanoidMultimodal behavior across a morphology familyEmbodiment-specific action targets or chunksTasks and related humanoid embodimentsBalance, contact and hardware variation
Latent or JEPA world modelPredictive state representationLatent future, value or planNew tasks with shared dynamics structurePlanning objective and action grounding
Generative world modelVisual and temporal generationFuture video or synthetic experienceData generation and policy evaluationPhysical fidelity and inference cost

Evaluating a foundation claim

Look for a matrix: rows vary tasks and environments, columns vary embodiment, adaptation data and evaluation shift. Report the checkpoint, every tuned component, the new robot data and a same-budget from-scratch baseline. VLAs near 95% on LIBERO drop below 30% under modest camera or initial-state perturbationsSource 26, and models above 90% fall to 0.0% under LIBERO-PRO's generalized settingSource 27.