A Comprehensive Review of Generative Physical Artificial Intelligence

In plain words

Robots have long followed fixed scripts: if this sensor fires, move there. This survey looks at a newer kind of robot brain, built from very large models that were first trained on huge amounts of images, text and robot recordings. Such a model can take a camera view and a spoken instruction and produce the motion that follows, and it can often handle objects or machines it never saw in training.

The authors sort these models into five groups. Some map what the robot sees straight to actions. Some generate a smooth motion plan by gradually cleaning up random noise. Some imitate human movement closely. Others predict how the world will change if the robot does something, which lets a robot try ideas in imagination or make training data. The survey explains each group, lists the tests people use to judge them, and discusses the hard problems: too little real-world data, the gap between simulators and reality, slow hardware, safety, and fairness.

The paper does not build or test a robot itself. It is a map of the field, and its numbers are taken from the works it cites.

TL;DR

• A survey of "Generative Physical AI" (GPAI): large generative models that directly drive or support physical systems such as robots and vehicles §1.
• Taxonomy of five families: Robot Foundation Models (RFM), Vision-Language-Action (VLA) models, Large Behavior Models (LBM), Diffusion Policy Models (DPM), and World Foundation Models (WFM) §1 fig. 2 tab. 3.
• Contains no experiments. Benchmarks and results are quoted from cited papers, and some quoted numbers have no citation (see Pitfalls).
• Useful as a reading map and a list of open problems. Not a source of implementation details.

Problem

Classical robot software is written for narrow tasks and breaks when parts, tools or obstacles change §1, Limitations of Early Robotic Systems. The survey argues that large pre-trained models can generalise across tasks, objects and robot bodies, but that the field lacks a shared structure for comparing them, benchmarks that predict real-world robustness, and safety evidence [§1, Related Surveys] §5.

Method

The paper's "method" is a classification and review, not an algorithm.

Components. A GPAI system is described as an embodiment with four linked parts: cognition and reasoning, perception, actuation, and a feedback loop. Data comes from demonstrations, teleoperation, simulation and digital twins §2 fig. 1. Learning paradigms covered are reinforcement learning (PPO, SAC, hierarchical RL, RLHF, RLAIF), supervised imitation, unsupervised and self-supervised learning §2, Learning Paradigms.

Taxonomy by function §3 fig. 2:

1. Generative control models link sensing to actuator commands. RFMs and VLAs belong here, with LBMs described as imitation-driven whole-body control §3, intro.
2. Generative action policy models generate trajectories. Diffusion policies are the example §3, intro §3.4.
3. Generative data models produce training and evaluation data. World foundation models are the example §3, intro §3.5.

Per-family summaries.

• RFM §3.1 fig. 3: a single large model over many robot datasets, with modality encoders, a fusion encoder and an action decoder. The paper contrasts task-specific models with general-purpose ones and gives PaLM-E and Gato as examples. Gato tokenises every modality into one vocabulary §3.1.
• VLA §3.2 fig. 4: a vision-language backbone plus an action decoder, trained end to end, so gradients from control reach the perception and language parts. Input streams are vision, language and proprioception. The survey walks through RT-1, RT-2, OpenVLA and Gemini Robotics VLA §3.2.
• LBM §3.3 fig. 5: a pre-training stage that uses reinforcement learning with two feedback signals (an authenticity signal from a discriminator and a physical-plausibility signal), and an inference stage with an embedding network and a policy network. Meta Motivo is the example, built on the FB-CPR method and using five networks: forward, actor, critic, backward and discriminator §3.3.
• DPM §3.4 fig. 6: a denoising model that turns Gaussian noise into an action sequence over a prediction horizon, conditioned on multimodal observations. Training inverts a noising process applied to expert trajectories, with epsilon-prediction or L2-type losses named as options. Receding-horizon execution runs only the first steps before replanning §3.4.
• WFM §3.5 fig. 7: a learned predictor of future observations given actions. Equation 1 gives the one-step form x_{t+1} = f(x_t, a_t) and equation 2 the H-step rollout over a candidate action sequence eq. 1 eq. 2. Autoregressive and diffusion generators are compared. NVIDIA Cosmos (Predict, Transfer, Reason) is the case study §3.5.

The planning loop that the survey describes for WFMs is sketched below.

# Receding-horizon planning with a world model  [§3.5] [eq. 1] [eq. 2]
# Input:  latent state x_t, goal g_t, world model f, planner
# Output: first action a_t to execute

def plan_step(x_t, goal):
    # Sample candidate action sequences of horizon H  [§3.5]
    candidates = planner.sample(horizon=H)
    best_score = -inf
    best_seq = None
    for each candidate in candidates:
        # Predict the H-step rollout under this candidate  [eq. 2]
        rollout = f(x_t, candidate)
        # Score the outcome: task reward minus safety penalties  [§3.5]
        score = reward(rollout, goal) - penalty(rollout)
        if score > best_score:
            best_score = score
            best_seq = candidate
    # Execute only the first action, then replan at the next step  [§3.5]
    return best_seq[0]

Notes on the sketch: the survey says several rollouts may be sampled per candidate to capture uncertainty, and that the choice of reward and penalty terms is left open §3.5. The sketch is the survey's description, not code.

Taxonomy synthesis. The paper maps families to stages of the control pipeline and notes a trend toward combining families, for example diffusion decoders inside VLAs §3.6 §3.4, Recent Advances.

Hyperparameters

The survey trains nothing. The table lists only model sizes and settings that the text states for the systems it discusses. Everything else is not stated.

Results

Author-reported, quoted from cited works. The survey itself runs no experiments, and the paper states that full empirical deployment is outside its scope tab. 1 §1, Table I notes. No baselines or datasets are set by the authors themselves.

Industry and market figures (robot counts, market sizes, surgical outcomes) appear in §1 and §4. They rely on press and analyst sources, not on measurements in the paper.

Paper vs code

No official code. The paper contains no repository, project page or code-availability statement. Searches for the title phrase and for "generative physical artificial intelligence" on GitHub returned no repositories.

Pitfalls

• No original experiments. The abstract speaks of "significant performance improvements". These are the results of the cited works, not of this paper §1, abstract. Do not read the survey as evidence that any single model works.
• Uncited numbers. The claim that VLAs with diffusion heads reach over 80% success against under 40% for vanilla VLAs has no citation in its paragraph §3.4, Recent Advances. The RT-1 67% figure over 700 tasks is also uncited in its sentence §3.2. Check the original paper before using either.
• Two RT-1 figures. The survey gives RT-1 67% over 700 tasks §3.2 and RT-1 32% on unseen tasks §4. They appear to measure different things. The paper does not say how they relate.
• Citation that may not support the claim. The Atlas "150 million simulations per maneuver" claim §4 cites the 2016 Atlas planning and control paper, whose title does not describe training at that scale. Verify the source before relying on it.
• Action-output mismatch for PaLM-E. RFMs are described as ending in an action decoder §3.1, architecture, yet PaLM-E is given as an example whose step-by-step output is text that needs external interpretation to become actions §3.1, exemplary models. Treat the RFM architecture figure as a generalisation, not a description of every example.
• Data formats. The data section says there are "two primary formats" and then lists three: structured, unstructured, and real-time streams [§2, Data Foundation].
• Diffusion notation. The denoising loop is written as "N-K steps" and as "K denoising iterations" in the same subsection, and figure 6 uses A_T^K for the noisy start §3.4 fig. 6. Use the original Diffusion Policy paper for step counts.
• Latency claims. The sub-50 ms requirement for fast manipulation is stated without a citation §5, Hardware and Real-Time Constraints.
• Autoregressive vs diffusion world models. The comparison §3.5 cites one work for diffusion and none for the autoregressive side.
• Taxonomy overlap. The paper names five families in its introduction but three functional roles in its taxonomy section §1 §3, intro. Where a family sits in the taxonomy (for example LBM under control, DPM under action policies) is the authors' choice, and other groupings are possible.
• Self-assessed coverage. Table I rates other surveys and this one for coverage, benchmarks and applications tab. 1. These marks are the authors' own judgement and were not checked.
• Figures. Figures 1 to 8 are described in the text only. They are not reproduced here.

Lineage

• Compared against: five surveys listed in tab. 1 on robotic manipulation with generative models, physics-aware generative vision, foundation models in robotics, world models for embodied AI, and robot learning with foundation models. Their content is not used on this card.
• Builds on: the individual models it describes, notably RT-1 and RT-2 §3.2, Diffusion Policy §3.4, Meta Motivo §3.3, and NVIDIA Cosmos §3.5. None of these has a card in this wiki yet.
• Superseded by: not known.