On 13 October Google published a separate post on applying Gemini 4 Argon to robotics. The model scored 87% on the ERQA benchmark, which tests spatial and physical reasoning. In one experiment it lifted a physical robot's success rate from 56% to 75%, using data it generated in a simulation it had built itself.
Key takeaways
- ERQA: 87% from pure reasoning — no other frontier model scores higher, according to Google
- Bimanual robot: success rate from 56% to 75%, task time from 35 s to 24 s
- Training data: over 4,500 episodes from a simulation built off a single RGB photo of the tabletop
- SO-101 arm: three self-improvement iterations, 58.9 s → 22.9 s → 17.4 s
- Antigravity and AI Studio restricted to the AI Ultra plan, API at $2 per million input tokens through 30 November
One model, four layers of the stack
Argon is not a VLA model driving joints directly. Google positions it as a general physical-reasoning model covering four areas: robot control, reconstructing simulations from recordings, data labeling and self-improvement.
In the control demo a Boston Dynamics Spot quadruped — 12 degrees of freedom in the legs plus a 6-DoF arm — learns to press a trash-bin foot pedal. Before touching real hardware, Argon builds an articulated MuJoCo model of the bin and practices three-leg balance inside it.
Google DeepMind's own clip: Argon driving a humanoid through a game of cornhole. This is not the FR3 experiment described below, but the same robot-control layer.
Simulation as a data factory
The strongest numbers concern data generation. From a single photo of a tabletop holding nine props, Argon reconstructed 3D meshes and collision shapes?collision shapes: Simplified geometry the physics engine uses for contacts. Cruder than the visual mesh, because it is evaluated hundreds of times per second., defined a curriculum of 77 single-arm and bimanual tasks, then collected over 4,500 episodes with ±10 cm pose randomization.
The input, task and episode counts and the final result are facts. The linear layout is a simplification — in Google's post reconstruction and verification run as an iterative loop. The diagram does not show which VLA model was fine-tuned, or how many iterations calibration took.
A VLA model fine-tuned on that data carried the gain over to a bimanual Franka FR3 robot: success rate from 56% to 75%, task time from 35 to 24 seconds. A separate real2sim?real2sim: Rebuilding a real scene and motion inside a simulator, so that further data can be collected virtually. path retargets a human hand trajectory from egocentric video onto a humanoid.
The humanoid in the real2sim path has far more axes than a lab arm:
Labeling and self-improvement
On a simple 6-DoF SO-101 arm, Argon cut pick-and-place time by reviewing its own logs and footage — from 58.9 s on the first iteration to 17.4 s on the third, … faster after two revisions.
Acting as an annotator, it split a 136-second recording of an ALOHA 2 robot folding origami into 6 semantic subtasks and 17 contact-level primitives, downweighting failed attempts to 0.15–0.25 for behavior cloning.
| Platform | Metric | Before | After |
|---|---|---|---|
| Franka FR3 (bimanual) | task success rate | 56% | 75% |
| Franka FR3 (bimanual) | task time | 35 s | 24 s |
| SO-101 (6-DoF) | pick-and-place time | 58.9 s | 17.4 s |
Why it matters
Real-hardware data was expensive, slow and hard to scale, so the pace of robotics was set by whoever owned the largest robot fleet. If a model can build the scene and produce the episodes itself, the edge shifts to reasoning quality rather than the number of arms on the floor. One caveat matters: the 19-percentage-point gain is a single Google measurement — and physical AI has neither a paper nor an independent replication to set against it.
What's next
- API pricing rises on 30 November from $2 to $4 per million input tokens and from $10 to $20 for output
- Google has not said whether Argon will reach Antigravity for plans below AI Ultra
- None of the robotics results has a paper or independent replication — Google's own post remains the only source
Sources
- Google Developers Blog — Gemini 4 Argon for Robotics
- Ars Technica — Gemini 4 Argon begins rolling out in Gemini app and Antigravity, but only if you pay
