Robotics in 2026: From Imitation to Self-Improving Physical Intelligence

Robotics is moving unusually fast right now, but much of the discussion mixes together concepts that belong at different layers.

Vision-Language-Action models. World Action Models. In-context learning. System 1/System 2. World models. Reinforcement learning. Self-improving robots.

They are not competing terms.

A more useful way to understand the current robotics stack is:

This distinction matters because the industry is progressing on all five axes simultaneously.

---

1/6 — Robot learning is becoming foundation-model learning

The old robotics recipe was usually:

task → collect task-specific data → train policy → deploy

That works well when the environment and task are narrow. It scales poorly when every new object, room, robot body or failure mode requires another data-collection cycle.

The new direction is a generalist robot foundation model: pretrain one policy across many tasks, robots and data sources, then reuse that prior for new situations.

Physical Intelligence (@physical_int)'s π0, for example, was trained with Internet-scale vision-language pretraining, Open X-Embodiment data and dexterous data collected across eight robot configurations. Its goal is explicitly to make specialization require less task-specific data. (physicalintelligence.company)

Figure (@Figure_robot)'s Helix follows the same generalist direction. A single model controls many household manipulation behaviors, with Figure reporting full upper-body control and handling of previously unseen objects without task-specific model weights. (figure.ai)

Google DeepMind (@GoogleDeepMind) has pushed this farther into whole-body robotics. Gemini Robotics 2, released in July 2026, uses the same model checkpoint across multiple embodiments and extends control from tabletop manipulation to walking, reaching, bending and dexterous manipulation. Its on-device model can be adapted to very different new robot bodies using a few hours of data, typically fewer than 200 examples. (deepmind.google)

This is why Sergey Levine (@svlevine) argues that robotics may follow a lesson already seen in language models: a sufficiently strong generalist can eventually outperform a narrow specialist because the generalist has a larger repertoire for handling exceptions. That is his research thesis rather than an established law, but it matches Physical Intelligence's development direction. (linkedin.com)

So the first major trend is straightforward:

robot intelligence is shifting from a collection of task policies toward reusable pretrained physical priors.

---

2/6 — VLA vs. WAM: two foundation-model recipes for robot learning

Most recent robot foundation models are Vision-Language-Action models, or VLAs.

Conceptually:

image + language + robot state → action

A pretrained vision-language model supplies semantic knowledge—objects, language, scenes and concepts—and an action component converts those representations into motor commands.

π0, Helix, Gemini Robotics and GR00T N1.x all belong broadly to this family. (physicalintelligence.company)

But a second foundation-model recipe is now becoming serious: World Action Models, or WAMs.

Instead of learning mainly from static image-language priors and mapping them directly to actions, WAMs learn something closer to:

current world + intended behavior → future world + actions

NVIDIA AI (@NVIDIAAI) describes WAMs as a second major robot-foundation-model recipe alongside VLAs. Its 2026 DreamZero model uses a pretrained video-diffusion backbone and jointly predicts future video and robot actions. (developer.nvidia.com)

The motivation is physical rather than semantic.

A VLM has seen enormous amounts of information about what things are. A video model has also been trained to predict how things change over time.

DreamZero's authors report more than 2× better generalization to unseen tasks and environments than the VLA baselines in their real-robot experiments. They also report transferring information from video-only demonstrations of humans or other robots, with more than 42% relative improvement on unseen-task performance from 10–20 minutes of video, and adaptation to a new robot embodiment using roughly 30 minutes of play data. (arxiv.org)

Jim Fan (@DrJimFan) described the idea publicly as a “second pre-training paradigm”: use video pixels as a shared physical representation across humans and different robot morphologies, rather than requiring every embodiment to share the same action space. (linkedin.com)

1X (@1x_tech) is independently exploring a related direction. Its 2026 1X World Model uses a 14B video model, 900 hours of egocentric human video, 70 hours of NEO embodiment data and an inverse-dynamics model trained on robot motion. The world model imagines a desired future video; the inverse model converts that transition into motor actions. (1x.tech)

But this should not be interpreted as “WAM replaces VLA.”

NVIDIA itself now discusses combining the strengths of both approaches. (linkedin.com)

And world models still have practical problems. 1X explicitly discusses visual, kinematic and physical inconsistencies in generated futures and uses inverse dynamics plus rejection mechanisms to ground them. Its January 2026 system also required a text prompt and was not yet fast enough for every reactive/contact-rich task. (1x.tech)

So a better taxonomy is:

VLA vs. WAM = foundation/pretraining architecture.

ICL = something different.

---

3/6 — ICL and memory let the same policy adapt while it is operating

In-context learning does not necessarily change model weights.

Instead:

experience enters the context → later actions condition on that experience

That distinction is especially important in robotics because many failures only become obvious after interacting with the physical world.

Physical Intelligence's 2026 Multi-scale Embodied Memory, or MEM, gives π0.6 both short-term visual memory and a compressed long-term language memory. The system tracks tasks lasting as long as 15 minutes and uses previous attempts to change its strategy. (pi.website)

One demonstration is deliberately simple but important.

When a robot cannot infer which direction a refrigerator door opens, a memoryless policy repeatedly makes the same mistake. With memory, it remembers the failed attempt and tries the opposite direction.

Likewise, after failing to grasp a flat chopstick, the memory-enabled policy modifies its next attempt rather than replaying the same motion. (pi.website)

That is in-context adaptation.

It is learning-like behavior, but the base model has not necessarily been retrained.

A second trend appears at the control level: companies are increasingly separating slow semantic reasoning from fast motor control.

Figure's Helix has a high-level VLM-based “System 2” running around 7–9 Hz and a reactive visuomotor “System 1” producing continuous control at 200 Hz. (figure.ai)

Physical Intelligence's Hi Robot similarly places a high-level VLM above π0. The upper model decomposes complex goals into smaller language-conditioned skills; the lower VLA executes them. (pi.website)

Google's current stack makes the split explicit too. Gemini Robotics ER 2 handles multi-step embodied reasoning and progress tracking, while Gemini Robotics 2 executes physical control. DeepMind reports that the reasoning layer can detect failed steps, replan and coordinate multiple robots across tasks lasting several minutes. (deepmind.google)

These systems differ technically, but the engineering pattern is converging:

reason at one timescale; control at another.

And again, this is separate from self-improvement.

A robot remembering that its previous grasp failed is adapting.

A robot using thousands of its own attempts to produce a better future policy is learning.

That leads to the more consequential trend.

---

4/6 — The data bottleneck is changing: from demonstrations to heterogeneous physical experience

Robot data is expensive.

Human teleoperation requires hardware, operators, resets and physical time. You cannot simply scrape trillions of robot trajectories from the Internet.

The industry response is not one new dataset. It is a change in what counts as robotics training data.

Physical Intelligence combines Internet vision-language data, data from many robot embodiments and task-specific robot experience. Its later human-to-robot work reports that human manipulation data produces additional gains on top of robot pretraining, with human data approaching target-robot data on some evaluated tasks. (pi.website)

Figure's Project Go-Big explicitly targets Internet-scale humanoid pretraining and reported zero-shot transfer from egocentric human video to robot navigation. (figure.ai)

1X uses egocentric human video before robot-specific grounding. (1x.tech)

DreamZero reports cross-embodiment learning from video produced by humans and other robots. (arxiv.org)

And NVIDIA Cosmos is building world-generation and physical-reasoning models specifically to create, transform and curate training environments and synthetic physical data. (developer.nvidia.com)

This changes the scaling equation.

Instead of:

more capability = more expert robot demonstrations

the frontier is increasingly trying to use:

**web/video pretraining

• heterogeneous robot data
• human egocentric video
• synthetic/simulated experience
• deployment rollouts**

Jim Fan's public framing is an “FSD-like” robotics data flywheel: increasingly capable policies make it possible to collect increasingly useful physical experience. He also argues, based on DreamZero, that diverse behavior may matter more than repeatedly collecting the same motion for world-action-model pretraining. (linkedin.com)

That last conclusion should not yet be generalized to every robot architecture. It is a result and hypothesis emerging from NVIDIA's WAM work, not a universal law of robot learning.

But the broader evidence is strong:

robotics training is becoming less dependent on clean, task-specific teleoperation alone.

---

5/6 — The most important shift: robots are beginning to learn from their own experience

This is where terminology needs to be strict.

“Generalization,” “self-correction,” and “self-improvement” are different capabilities.

Level 1 — Generalization

The robot encounters something it was not explicitly trained on and succeeds because of prior knowledge.

No learning occurs during deployment.

DreamZero zero-shot tasks and π0.5 operating in previously unseen homes are examples. (arxiv.org)

Level 2 — In-context adaptation

The robot remembers what happened during the current task and changes its next attempt.

The underlying policy weights need not change.

Physical Intelligence's MEM refrigerator and chopstick examples belong here. (pi.website)

Level 3 — Self-improvement from autonomous experience

Now deployment actually creates training data that improves subsequent behavior.

This is already real.

Google DeepMind's RoboCat demonstrated an early version in 2023. After receiving 100–1,000 demonstrations for a new task or robot, a specialized policy would practice the task roughly 10,000 times, and those self-generated trajectories were incorporated into the next RoboCat training cycle. (deepmind.google)

Physical Intelligence's π*0.6/RECAP moves this idea into modern VLA training. RECAP combines demonstrations, expert corrections and autonomous robot experience. Physical Intelligence reports that on some difficult tasks this more than doubled throughput and roughly halved failure rates. (arxiv.org)

Its newer RLT work makes the loop much faster. Instead of retraining the entire VLA, a small RL actor and critic learn to modify the VLA's actions from deployment experience. PI reports improvement with minutes to hours of robot data—including 15 minutes of actual data for one precision task—and up to 3× faster execution on the evaluated fine-manipulation stages. (pi.website)

This is much closer to the important long-term idea:

the robot does not only consume a static demonstration dataset; its own failures and successes become part of the learning process.

---

6/6 — But “self-evolving robot” should currently be used carefully

There is an even stronger version of the idea:

Give a robot a body and an environment.
Let it explore, discover skills, evaluate itself, practice and improve—without requiring humans to explicitly demonstrate every behavior or continually tell it what action to perform.

Pieces of that system now exist.

NVIDIA's 2026 ASPIRE is explicitly described as a self-improving continual-learning system. It observes rollout traces, diagnoses failures, rewrites robot-control programs, validates the repair and stores successful solutions in a growing skill library. It also uses evolutionary search to generate different task sequences and control programs instead of optimizing only one trajectory. (research.nvidia.com)

That is genuine autonomous skill refinement.

But there is an important boundary.

NVIDIA explicitly states that ASPIRE is not yet a fully autonomous real-world learner. Its current system is mainly demonstrated in simulation, relies on a frontier LLM, uses a predefined primitive API, and still needs better real-world success detection, safe resets, calibration and safety monitoring. (research.nvidia.com)

1X makes an equally ambitious argument for its world-model approach. The company says NEO's experience can eventually form a flywheel of exploration, evaluation and policy refinement instead of relying on expert demonstrations. But its January 2026 article describes this under “towards full world model autonomy”; the demonstrated system still receives a text prompt at inference and still requires substantial embodiment and inverse-dynamics training. (1x.tech)

So among the frontier systems reviewed here, there is not yet evidence for a general-purpose physical robot that autonomously invents arbitrary useful goals, learns them from scratch in the open world, continually changes itself, and requires no human demonstrations, prompts, reward design or safety infrastructure.

What is supported by evidence is more interesting than the usual “humanoid demo” narrative:

Robots are gradually moving from:

demonstrate → imitate → deploy

toward:

pretrain → attempt → observe → remember → evaluate → practice → improve

That is the trend I would watch most closely.

WAMs matter because they may provide stronger priors about how the physical world changes.

VLAs matter because they bring Internet-scale semantic knowledge into action.

Memory and ICL matter because a deployed robot cannot afford to repeat every mistake.

Hierarchical reasoning matters because “clean the kitchen” cannot be executed at the same timescale as fingertip control.

Human video, world models and simulation matter because robot demonstrations alone are too expensive to become the Internet of physical intelligence.

And reinforcement learning from deployment matters because demonstrations place a ceiling on performance: eventually the robot has to learn from states created by its own behavior, including states where it made mistakes.

The clearest description of the current robotics transition is therefore not:

VLA → WAM

or

WAM → ICL

It is:

static imitation → generalist physical priors → adaptive execution → experience-driven self-improvement.

The architecture underneath that progression may ultimately contain VLAs, WAMs, memory, reasoning models and reinforcement learning at the same time.

That is why the most consequential robotics race may not be about who produces the most impressive pretrained robot.

It may be about who closes the learning loop after deployment.

#Robotics #PhysicalAI #EmbodiedAI #RobotLearning #VLA #WorldModels #HumanoidRobotics