1. Comprehensive Hardcore Comparison Table of the Ultimate Multimodal Robot Model (VLA / VOA)

2. 12 Core Models: Precise Text Analysis

To make readers or collaborators understand, we need to transform these cold structures into their “personalities” in academia and industry.

🏛️ The Era of Founding and Establishment: Google’s Dominance

• RT-1: The pioneer of the VLA era It proved something to the world—there is no need to design complex control algorithms specifically for robots. By mapping the actions of the robotic arm into “words” (discrete tokens) in a text dictionary and feeding them into a Transformer for computation, robots can learn to perform tasks.
• RT-2: A massive system that delivers a low-dimensional attack It addresses the issue of robots lacking common sense. Google aggressively integrated a multimodal language model with 55 billion parameters directly into robots. This enables robots not only to understand commands but also to develop “thought chains” similar to humans. For example, they can understand the instruction “throw away expired food” and automatically identify rotten bananas from the table.
• RT-X: A “world-wide unified” system for cross-domain data Strictly speaking, it is not a new architecture, but rather an “ultimate weight and dataset” created by Google in collaboration with top global academic institutions. It solves the problem where laboratories train robots using only their own data, resulting in the robots not responding when switched to a different robotic arm. This enables true Cross-Embodiment control.

🐙 Era of Open Source Explosion: The Celebration in Academia

• Octo: The purest VOA action robot This is a model developed by the Berkeley and Stanford teams specifically for robots that do not require dialogue. It has no bulky language large model overhead; instead, it relies on a highly lightweight Transformer and diffusion model to perform actions based solely on visual input. In pure imitation learning tasks, it is the fastest-performing open-source benchmark.
• OpenVLA: The versatile warrior of the open-source community It directly addresses the closed-source and overly complex nature of RT-2. Based on a 7B-level Llama 2 model, it integrates a hybrid visual foundation capable of understanding semantic and spatial information. It is currently the preferred “out-of-the-box” robot brain for most startups and research labs worldwide.
• RoboFlamingo: The multi-frame memory expert that cures amnesia Previous models used “single-frame input” (seeing a single image and performing an action), which often caused robots to forget their initial goal when handling long-term tasks. RoboFlamingo incorporates a dedicated explicit policy head at the end of the network, ensuring it remembers historical multiple frames, making it extremely stable for handling long-term tasks like “opening a drawer to pick up an apple and then closing it.”

🚀 Era of Extreme Evolution: Solving Specific Diseases

• SpatialVLA: A 3D Dominator Breaking Through 2D Limits Solves the fundamental problem of traditional models being “flat in perspective.” It forcibly injects Ego3D deep positional encoding into the network and introduces an adaptive spatial action mesh. When the robotic arm performs insertion and precise grasping tasks requiring high spatial accuracy, its success rate achieves a dramatic improvement.
• SmolVLA: A Ultra-Lightweight, Accessible Power Tool A miracle created by Hugging Face for devices with limited computing power. Through brutal network layer jumps and feature compression, it tightly limits the number of parameters to 450M. It completely resolves the issue that VLA models rely on hundreds of thousands of server-grade graphics cards, allowing ordinary people to train a household-sized robotic arm using just a laptop.
• SwiftVLA: A Temporal Spatiotemporal Prediction Expert When the target object is moving (e.g., throwing and catching balls, grasping items on conveyors), the static VLA fails completely. This model innovatively introduces a 4D streamVGGT spatiotemporal feature and mask reconstruction strategy. With only 450M parameters, it enables large models to “predict” the trajectory of the robotic arm and object in the future. It offers strong adaptability to dynamic scenarios and outperforms large models in inference speed.
• StarVLA: A Co-training Arsenal for Image and Text Action Collaboration It provides an elegant underlying training logic, enabling developers to seamlessly co-train massive amounts of internet text data with expensive robotic arm action data in the same framework, significantly reducing the engineering barrier for developing general-purpose robots.

🟢 Industrial Harvesting Era: NVIDIA’s Dimension Reduction Reconfiguration

• GR00T N1.5: The core strength of cross-body movement fed by synthetic data A industrial-grade nuclear weapon used by NVIDIA to showcase its capabilities. Its biggest innovation lies in no longer relying heavily on real data collected by human-controlled robotic arms, but instead absorbing large amounts of “synthetic data” automatically generated by NVIDIA Omniverse and GR00T-Dreams within the simulation engine. It is equipped with a highly smooth 16-layer Flow Matching Transformer at the backend, resulting in delicate and extremely stable motion execution.
• GR00T N1.6: The ultimate evolved form with deep physical logic The current pinnacle of humanoid robot technology. Its language foundation has been replaced directly by NVIDIA’s Cosmos Reason reasoning large model designed for the physical world. This grants it a “System 2 (slow thinking)” ability similar to that of humans—when given extremely vague instructions, it first conducts mental simulations based on physical常识, then passes the motion intent to a motion prediction head with 32 layers of depth, ultimately transforming it into complex, coordinated movements for the humanoid robot.

🏠 Industrial Implementation and Infrastructure Era: Horizon (Horizon Robotics)'s Full-Stack Breakthrough

HoloBrain-0: A “All-round Manager” with a Physical Structure Understanding It effectively solves the problem of poor adaptation of VLA models when transferred between different robot hardware. Unlike other models that require robots to memorize the mapping between pixels and movements, this model explicitly uses the robot’s “muscles and bones” (URDF dynamics description) and “eye positions” (camera internal and external parameters) as input. With the help of a powerful “spatial enhancer,” it can plan directly in 3D coordinate systems, achieving strong cross-robot compatibility.

What’s even more advanced is the accompanying RoboOrchard infrastructure. Through a “test-driven” closed loop iterative data strategy, it no longer blindly piles up massive amounts of data, but instead precisely “compensates” for failed cases. This enables its ultra-lightweight version with 0.2B parameters to smoothly complete challenging long-term tasks such as folding clothes and stacking cardboard boxes on domestic edge chips costing just a few hundred yuan, using only about 30 hours of real data.