Industrial robots are extraordinarily capable inside a carefully designed routine. Move the part, change the lighting, alter the gripper, or ask for a new task, and that competence can disappear. Robotics foundation models aim to loosen that constraint by learning reusable patterns across many objects, instructions, environments, and robot bodies.
The emerging architecture is often called vision-language-action, or VLA. A model receives camera images and a natural-language instruction, then produces actions a robot can execute. Instead of engineering every motion from scratch, developers start with a broad model and adapt it to a machine and workplace. The ambition resembles language models: pretrain once on diverse data, then reuse the learned representation across many tasks. Physics makes the analogy imperfect.
A shared model for perception and action
The field’s data strategy changed with Open X-Embodiment, a collaboration that assembled more than one million real robot trajectories across 22 embodiments, from single arms to bimanual systems and quadrupeds (Open X-Embodiment). Pooling data allows a model to encounter more objects and behaviors than any one laboratory could collect. It also tests whether knowledge learned on one robot can help another.
Google DeepMind’s 2025 Gemini Robotics model extended a multimodal model with physical actions as an output. The company reported more than double the performance of other state-of-the-art VLAs on its generalization benchmark and demonstrated adaptation across ALOHA, Franka, and Apollo platforms (Google DeepMind). Its on-device version could be adapted to new tasks with 50 to 100 demonstrations, according to the company, while running locally to reduce latency and dependence on connectivity (Google DeepMind).
By July 2026, Gemini Robotics 2 was targeting whole-body humanoid control, longer sequences, and collaboration among multiple robots. DeepMind described tasks lasting several minutes and involving hundreds of decisions, alongside a safety benchmark for deciding when to refuse an unsafe tool call or request human intervention (Google DeepMind). These are developer-reported evaluations, not proof of unrestricted autonomy, but they show how quickly the target is expanding.
NVIDIA has pursued a parallel route. GR00T N1, released in March 2025 as an open, customizable humanoid foundation model, combines a vision-language component for reasoning with a diffusion-based component for continuous action generation (NVIDIA Research). Open weights and simulation tools can lower the cost of experimentation, while synthetic data can supplement scarce demonstrations.
Physical data changes the economics
Internet text is abundant and easy to copy. Robot experience is expensive. It requires hardware, safe space, maintenance, human demonstrations, and recovery from failures. Each body has different joints, sensors, payloads, and control rates. Even the meaning of a simple action such as grasp depends on friction, geometry, and force.
Simulation helps generate rare conditions and large volumes, but the simulation-to-reality gap remains. A model trained on idealized contact may fail on a reflective object, deformable bag, sticky drawer, or worn floor. Real-world data is therefore still needed to calibrate what the simulator misses.
Foundation models may nevertheless alter deployment economics. A warehouse or laboratory could fine-tune a general policy instead of commissioning a separate stack for every manipulation task. Natural-language instructions could make reconfiguration easier for non-specialists. The strongest near-term case is not a robot that can do everything; it is a reusable starting point that reduces engineering for a bounded family of tasks.
Generality raises the safety burden
A conventional robot cell is safe partly because its behavior is narrow and predictable. A model that interprets open-ended instructions creates more possible states. It can misunderstand an object, choose an unsafe path, or continue when it should stop. As DeepMind notes, in the physical world “surprises are the norm”.
High-level learned policies should therefore sit above deterministic safety controls: collision avoidance, speed and force limits, emergency stops, geofencing, and body-specific stability systems. Deployers need task-level tests, logs, uncertainty thresholds, change control, and a defined handoff to a person. NIST’s human-robot work emphasizes repeatable metrics for effective teaming and the need to measure shared situational awareness, not just whether a robot eventually completes a task (NIST).
The decisive benchmark will be sustained performance outside curated demonstrations: unfamiliar objects, changing environments, long task chains, diverse users, and safe recovery after error. Robotics foundation models are making machines more adaptable. General skill will require something harder—evidence that adaptation remains dependable when the world refuses to stay still.