Gemini Robotics 1.5
Google DeepMind, US · September 2025
- Parameters
- not published
- No parameter count published.
- GPU memory
- not published
- weights at bf16, computed
- Inference latency
- not published
- per action step
- Weights
- closed
- no public checkpoint
What it is
Gemini Robotics 1.5 is the thinking VLA of the Gemini family, built on the Gemini 2.5 generation. Two ideas carry it: a Motion Transfer mechanism that lets one model learn from heterogeneous multi-embodiment data, and an interleaved natural-language reasoning process so the robot thinks before acting. It controls ALOHA, bi-arm Franka and the Apollo humanoid from a single checkpoint with no robot-specific post-training, and shows zero-shot skill transfer between them. The report presents the VLA evaluations as figures rather than tables, so precise success rates are not publicly quotable, and the model itself was never released beyond selected partners.
Architecture
- Backbone
- Built on the latest generation of Gemini at the time of writing, cited in the report as the Gemini 2.5 model family.
- Action head
- Thinking VLA. It interleaves motor actions with a multi-level internal reasoning process in natural language, so the model produces reasoning text before and between action segments. A Motion Transfer mechanism aligns action spaces across embodiments.
- Parameters
- No parameter count published.
- Pretraining data
- Heterogeneous multi-embodiment robot data spanning ALOHA, bi-arm Franka and the Apollo humanoid, combined through the Motion Transfer training recipe, on top of Gemini pretraining. The report does not give a corpus size in hours or episodes.
- Embodiments
- ALOHA, Bi-arm Franka, Apptronik Apollo humanoid
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
No results in this category are published for this model.
Fine tuned tasks
No results in this category are published for this model.
On real hardware
Controls ALOHA, bi-arm Franka and the Apollo humanoid from one checkpoint without robot-specific post-training. The report documents zero-shot skill transfer between embodiments, where ALOHA performs tasks whose training data was collected only on the bi-arm Franka and vice versa, and the humanoid performs skills present only in ALOHA data. Results for the VLA are presented as figures rather than numeric tables, so no exact per-task success rates are quoted here.
Fine tuning it yourself
No public fine-tuning path. Access was limited to selected partners and a trusted tester program.
Where it helps, where it does not
Strengths
- One checkpoint controls three very different robots with no robot-specific post-training
- Motion Transfer produces demonstrated zero-shot skill transfer across embodiments, measured as success rate and not only partial progress
- Thinking before acting improves decomposition of multi-step tasks and makes behavior interpretable
- Largest benefit from cross-embodiment data appears exactly where data is scarcest, on the humanoid
Limits
- Closed weights, available only to select partners, never released publicly
- The report publishes VLA results as figures without a numeric results table, so success rates cannot be quoted precisely
- Motion Transfer helps least on the humanoid, where the embodiment gap is largest, which is the case that most needs it
- No latency, control frequency, parameter count or hardware requirement published
- Superseded by Gemini Robotics 2 in July 2026
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2510.03342
- https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/
Entry last checked 2026-08-11.
Try a policy on a real arm
A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.