Executive Summary

To understand Vision-Language-Action (VLA) models, it helps to see what came before. Traditional robot software treated three stages as separate modules: a perception module to recognize objects from camera images, a planning module to compute paths, and a control module to move joints — each developed independently and governed by hand-written rules. A VLA model collapses all three into a single neural network, enabling end-to-end robot control through natural language commands.

The Three I/O Channels of a VLA

• V Vision — perceives the surroundings via camera images. "The plate is on the left shelf."
• L Language — understands natural language instructions. "Move the plate to the right shelf."
• A Action — generates joint torque/position control signals. "Elbow 45°, wrist 20°..."

VLA vs. World Model — Drawing the Line

A concept often confused with VLA is the world model. The two serve different purposes. A VLA maps observations directly to actions; a world model predicts future states. In 2025, however, the boundary is blurring. NVIDIA's Cosmos World Foundation Model has been absorbed into GR00T N1.7's VLM backbone. We explore the implications of this convergence in Section 7.

Why Now?

VLA models surged in 2024–2025 because three prerequisites converged simultaneously. First, internet-scale VLMs (Vision-Language Models) became strong enough to jointly process language and vision. Second, advances in teleoperation and synthetic data generation drove down the cost of collecting robot data. Third, humanoid hardware from Figure, 1X, Unitree, and Apptronik proliferated rapidly, creating explosive demand for general-purpose policies.

NVIDIA's Isaac GR00T series has evolved from the original N1 release in March 2025 through N1.5 and N1.6 to N1.7. The core hypothesis has remained consistent: "Just as humans deliberate in novel situations and act reflexively in familiar ones, robots should do the same." This intuition materializes as the Dual-System (Action Cascade) architecture.

Version Lineage

Action Cascade — A Brain in Two Parts

The core of GR00T N1.7 is the separation of two systems. System 2 (slow thinking) formulates a plan, then System 1 (fast reflexes) moves the body. An MLP connector bridges the two, and the quality of this connector largely determines overall performance.

[Input: Camera images + natural language command]
        ↓
┌——————————————————┐
│  System 2: Slow Thinking (VLM)        │
│  Cosmos-Reason2-2B (Qwen3-VL based)   │
│  Native aspect ratio, no padding       │
│  → Plans "where to place the plate"    │
└—————————│————————┘
               ↓ MLP connector (plan tokens)
┌——————————————————┐
│  System 1: Fast Reflexes (DiT)         │
│  Diffusion Transformer                 │
│  Flow Matching training                 │
│  → Real-time "bend wrist by N°"       │
└—————————│————————┘
               ↓
[Output: Continuous joint position/torque signals]

EgoScale — Learning from Human Eyes

GR00T N1.7's most original contribution is EgoScale. Instead of robot data, it learns from first-person (egocentric) videos filmed by humans. Cooking at home, assembling parts in a factory, treating patients in a hospital — every hand movement captured from a human perspective becomes training data.

• • 20,854 hours of human egocentric video
• • 20+ task categories — manufacturing, retail, healthcare, household, and more
• • First discovery of a manipulation scaling law — scaling data from 1,000 to 20,000 hours more than doubled average task completion rates
• • Isaac Lab/Isaac Sim generates unlimited synthetic data to fill remaining gaps

Google DeepMind unveiled Gemini Robotics in March 2025, built on Gemini 2.0. Where NVIDIA splits the brain in two, Google keeps everything inside a single large model but processes "thinking" and "acting" sequentially. Version 1.5 introduces two critical innovations: Think-then-Act, where the model reasons in natural language before producing motor commands, and Motion Transfer, which unifies data from multiple robot platforms.

Two Variants: Robotics vs. Robotics-ER

Architecture — Language First, Action Second

[Input: Camera images + natural language command]
        ↓
┌——————————————————┐
│  Gemini 2.0 Multimodal Model           │
│  Vision encoder + language encoder      │
│  → Internal reasoning chain (1.5 new)  │
│    "First check plate position,          │
│     then extend the arm..."              │
└—————————│————————┘
               ↓ Action tokens + embeddings
┌——————————————————┐
│  Motion Transfer (MT) Layer             │
│  Heterogeneous platforms → unified     │
│  ALOHA / Bi-arm Franka / Apollo etc.    │
└—————————│————————┘
               ↓
[Output: Platform-specific joint control commands]

Motion Transfer — One Model, Many Robots

One of the hardest problems in robotics is that every robot has a different body. A policy trained on the ALOHA dual-arm robot cannot run on a Franka arm out of the box — different arm lengths, joint counts, and degrees of freedom (DOF) make direct transfer impossible.

Motion Transfer tackles this head-on. It projects motion data from multiple robots into a single shared physical representation space, enabling a single GR 1.5 checkpoint to control ALOHA, Bi-arm Franka, and Apollo humanoids alike.

Physical Intelligence (PI) is a Silicon Valley startup, but its founding team is anything but ordinary — roboticists from Google Brain, DeepMind, Stanford, and UC Berkeley. Their technical bet diverges from both NVIDIA and Google: remain hardware-agnostic and push past the limits of dexterous manipulation by "generating actions via diffusion."

Version Lineage

• π0 2024.10 — First VLA Flow model. PaliGemma 3B + Flow Matching. Dexterous manipulation focus.
• π0-FAST Faster inference variant optimized for real-time control.
• π0.5 Current 2025 — Enhanced open-world generalization. Released openpi as open source. Provides DROID dataset training pipeline.

Architecture — Diffusion on Top of PaliGemma

[Input: Camera images + natural language command]
        ↓
┌——————————————————┐
│  PaliGemma 3B (VLM backbone)            │
│  Internet-scale image-text pre-training  │
│  Scene understanding + language semantics│
│  Embeddings aligned with language tokens │
└—————————│————————┘
               ↓ Multimodal embeddings
┌——————————————————┐
│  Flow Matching Head (diffusion-based)    │
│  Learns path from noise → target action │
│  Produces continuous high-freq sequences │
└—————————│————————┘
               ↓
[Output: Continuous joint control signals (high-freq)]

Flow Matching — Why Diffusion?

Classical diffusion models (DDPM) generate data by traversing thousands of steps from noise to target. Flow Matching simplifies this process by learning a more direct "flow path" from noise to target, producing high-quality continuous actions in fewer steps.

This matters because of the nature of robot manipulation. "Pick up the plate and place it on the shelf" requires dozens of continuous, high-frequency motions. Models that represent actions as discrete tokens struggle to capture such smooth trajectories. Flow Matching learns action distributions directly in continuous space, giving it an edge in dexterous manipulation.

Signature Tasks and Real-World Performance

The tasks PI chose for its demos are telling. Folding laundry, loading a dishwasher, setting a table — all require irregular shapes and delicate force control, making them quintessential dexterous manipulation challenges. This positioning differs from GR00T's "whole-body humanoid control" and Gemini's "230 diverse tasks."

In zero-shot evaluation (no fine-tuning), π0.5 achieved an average 42.3% task progress rate. The number may look modest, but it represents a meaningful improvement over prior work. Put another way, about six out of ten attempts still fail — a candor that is characteristic of PI's research style.

What picture emerges when we line up all three models against the same yardstick? The table below compares them across 13 dimensions spanning architecture, data, ecosystem, and openness. The goal is not to declare a winner, but to illuminate the different problems each team chose to solve and how they solved them.

The data strategy behind each model matters as much as the architecture itself. "What kind of data should you train on to build a good robot policy?" The three teams gave completely different answers. This is not merely a technical choice — it reflects a worldview about which data best captures the essence of robotic behavior.

No discussion of VLA architectures is complete without addressing world models. Should a robot "imagine the consequences of its actions before acting," or should it "predict the best action directly from the current observation"? This philosophical question has profoundly shaped each team's technical choices.

NVIDIA Cosmos — Folding the World Model into the VLA

NVIDIA gave the most emphatic answer to this question. Cosmos is NVIDIA's World Foundation Model platform. Cosmos-Predict forecasts future video, while Cosmos-Reason produces physics-grounded decision-making via natural language chains.

GR00T N1.5 co-trained flow matching with a world modeling objective. N1.7 went further, adopting the Cosmos-Reason model outright as its VLM backbone. NVIDIA's message is clear: "The world model must live inside the VLA."

Three Teams, Three Hypotheses