**RT-2 (Robotics Transformer 2)** is a breakthrough Vision-Language-Action model announced by Google DeepMind in July 2023 (paper 'RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control', Brohan et al., arXiv:2307.15818). The successor to RT-1, but unlike RT-1 (which was a specialized 35M-parameter transformer trained exclusively on robotic data), RT-2 **co-trains** a massive pretrained Vision-Language Model (PaLI-X 55B or PaLM-E 12B) on a **mixed dataset** of web data (~10B image-text pairs) + robotic data from RT-1 (~130k episodes).
The key innovation: **actions as tokens**. RT-2 discretizes the robot's action space (6-DOF end-effector translation/rotation + gripper) into 256 bins per dimension and treats them as **additional tokens in the VLM's vocabulary**. This lets the model uniformly generate the 'answer' as a sequence of tokens — whether textual (VQA) or a robot action. This approach enables **emergent capabilities**: RT-2 can perform semantic tasks requiring chain-of-thought reasoning ('Move banana to the sum of two and one' → count the items, find '3' as the answer, approach the third object) that are absent from the training data.
Results: RT-2 achieves **62% success in generalization scenarios** (novel objects, novel instructions, novel backgrounds) vs. 32% for RT-1. Experiments were run on Everyday Robots mobile manipulators (a Google internal project — discontinued since 2023) and Franka. RT-2 **is not open source** — Google DeepMind only released a checkpoint of a small replication version, with no full PaLI-X/PaLM-E weights. Successors: **RT-X / Open X-Embodiment** (October 2023, cross-embodiment generalization), **Gemini Robotics** (March 2025, Apptronik Apollo integration).
RT-2 launched the era of 'foundation models for robotics' and influenced an entire generation of subsequent models: OpenVLA (Stanford/Berkeley, open-source replication), π0 (Physical Intelligence), Octo (Berkeley), CogACT. Most VLAs from 2024-2026 inherit the 'tokens as actions' architecture from RT-2.
A Runtime is the environment or execution layer used to run code, load libraries, manage dependencies, and operate applications or services — either in real time or during normal system operation. In robotics this includes real-time operating system (RTOS) runtimes, ROS 2 executor runtimes, containerised execution environments (Docker, podman), and embedded C++ runtimes on microcontrollers.
Internal Google DeepMind experiments on Everyday Robots mobile manipulators (a discontinued project) — 6,000+ eval scenarios in 2023. RT-2 became the foundation of the Open X-Embodiment program (X-Embodiment / RT-X, October 2023, 21 institutions + Google DeepMind, 1M demonstrations from 22 robotic platforms). The successor **Gemini Robotics** (March 2025) is integrated with Apptronik Apollo and Boston Dynamics Spot.
No public weights repository. The arXiv:2307.15818 paper has 1,500+ citations (Q1 2026). The robotics-transformer2.github.io project site with video demonstrations. RT-X / Open X-Embodiment github.com/openx-embodiment ~2.8k★. The #RT-2 / #VLA hashtag on X/Twitter generates ~50-100 mentions/day.
Boston Dynamics bipedal humanoid robot. The fully electric generation unveiled in 2024 succeeds the hydraulic Atlas that was retired after more than a decade of research.
Figure 03 is the third-generation humanoid robot from Figure AI, designed for Helix, home environments, and scalable mass production.
Bipedal humanoid robot by Unitree Robotics, designed as a compact research, development, and developer platform.
Ubuntu 24.04 LTS 'Noble Numbat' — supported until April 2029. The host for ROS 2 Jazzy.
RT-2 was not released as open weights. Google has only published the paper and inference code for a small replication version. The community built OpenVLA (Stanford/Berkeley) as an open-source successor with an architecture similar to RT-2.
License family: Proprietary – Commercial
Successor to RT-2 based on Gemini 2.0 — a multimodal VLA with an Apptronik (Apollo) partnership.
Derived architectures — AutoRT (scaling robot data collection) + SARA-RT (efficient self-attention).
RT-2 trained on the expanded Open X-Embodiment dataset.
Cross-embodiment training: 1M demonstrations across 22 robots + 21 institutions. First cross-embodiment generalization benchmark.
RT-2 — co-training PaLI-X 55B with robotic data. Introduces 'tokens as actions'. Emergent reasoning capabilities.
Robotics Transformer 1 — 35M parameters, trained exclusively on robotic data (130k Everyday Robots episodes). The baseline benchmark.
