Gemini Robotics 2: Advancing Whole-Body Intelligence and Dexterity

Gemini Robotics 2 enables whole-body coordination and multi-robot teamwork

Google DeepMind has released Gemini Robotics 2, an intelligence layer designed to move robotics beyond pre-programmed, repetitive tasks toward general-purpose physical AI. The system enables robots to reason through complex movements, coordinate their entire bodies—from feet to fingertips—and collaborate with other robots to complete multi-step workflows.

The Three-Model Architecture

Gemini Robotics 2 is powered by three specialized models that divide the labor between high-level reasoning and low-level motor control:

  • Gemini Robotics 2 (VLA): A Vision-Language-Action model that converts visual and linguistic inputs directly into motor control. It manages full humanoid movements and high-dexterity manipulation for both five-fingered hands and parallel grippers.
  • Gemini Robotics ER 2 (ER): An Embodied Reasoning model (VLM) that acts as the agent's "brain." It handles human communication, environmental understanding, and the planning of multi-step tasks lasting several minutes. This model also coordinates multi-robot collaboration.
  • Gemini Robotics On-Device 2 (VLA): An efficient VLA model optimized for local execution to eliminate network latency. It features "motion transfer" capabilities, allowing it to adapt to new robot embodiments (different shapes, sensors, and degrees of freedom) within a few hours using fewer than 200 examples.

Whole-body control and advanced dexterity

Gemini Robotics 2 expands physical AI from simple table-top tasks to full-body coordination. By translating intent into whole-body control, the system can manage complex sequences such as walking to a location, crouching, and precisely placing an object on a bottom shelf.

High-Precision Manipulation

The system demonstrates a significant leap in dexterity across different hardware end-effectors:

  • Five-fingered hands: Using the 22 degree-of-freedom SharpaWave hand on the Apollo 2 humanoid, the model can perform delicate tasks like tying knots or sealing ziplock bags.
  • Parallel grippers: On the Franka Duo platform, the model executes complex packing tasks using standard two-fingered grippers.

Agentic reasoning and safety orchestration

To manage tasks that span several minutes and involve hundreds of decisions, Gemini Robotics ER 2 provides high-level orchestration. It observes the environment, reasons through the necessary steps, and self-corrects if a specific action fails.

Safety and the ASIMOV-Agentic Benchmark

Google has introduced ASIMOV-Agentic, a new benchmark for agentic safety orchestration. This framework measures a robot's ability to:

  1. Refuse unsafe tool calls requested by a VLA.
  2. Predict whether a task is physically possible.
  3. Proactively request human intervention when uncertainty is high.

Additionally, Gemini Robotics ER 2 is designed to detect human proximity and trigger safety stops to comply with collaborative safety standards.

Technical Perspectives and Community Discussion

While the announcement highlights significant progress, the technical community has raised several critical points regarding the practical deployment of LLM-based robotics:

Latency and Control

Some developers argue that using a full LLM for actuation may introduce unacceptable latency (estimated at 1-2 seconds), which is impractical for real-time robotics. A common counter-proposal is to use ML for vision but maintain traditional PID-style linear/nonlinear control for motion actuation.

Hardware and Safety Concerns

Discussion among users highlighted the "lawsuit risk" of powerful humanoid robots in home environments, noting that a robot strong enough to perform heavy labor is also strong enough to cause serious injury if it malfunctions.

Performance Metrics

Critics have pointed to accompanying technical data suggesting success rates of approximately 60% and accuracy of 80%, arguing that these figures indicate the technology is not yet production-ready for autonomous household or industrial use.

"If progress is as fast as LLMs, this could have massive applications in a few years." — @FartyMcFarter

"The big one will be using AI to design and build highly specialized robots to automate mining and manufacturing." — @baron816

Sources