Google and NVIDIA make robot learning more adaptable
Gemini Robotics 2.5 is reaching Google’s partners as NVIDIA presents GR00T 2’s few-demonstration learning; both advances remain task-specific, provider-reported results.
The useful robotics question is how much new training a machine needs when the task changes.
Google’s November 10 Gemini Robotics 2.5 release broadens manipulation across more than 200 evaluated tasks and begins rolling out to partners. NVIDIA’s November 9 GR00T 2 announcement attacks the same bottleneck through learning from human video and demonstrations on other robot bodies. [1][2]
NVIDIA reports 40% average trial success across 50 tasks new to the target robot, rising to 71% after one demonstration per task and 95% after five. The separate statement that 40 of 50 tasks succeeded at least once measures task coverage, not an 80% reliable robot. Google reports 96% success in a Mercedes-Benz automotive-kitting pilot with Apptronik’s Apollo 2; this is a different evaluation and cannot rank the two models. [1][3]
For a business, the near-term opportunity is a repeatable, bounded task whose failure cost is understood. Ask for trial-level reliability, recovery behavior and performance after objects or layouts change. Google’s release is partner access; GR00T 2’s project page says code is coming in December. Neither announcement establishes that a general-purpose robot can be bought and deployed today. [1][3]
Sources & details 4 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Google DeepMind — Gemini Robotics 2.5
Partner rollout, not general public robot API. Authors evaluated over 200 tasks and report 96% automotive kitting pilot success at Mercedes-Benz using Apptronik Apollo 2. No trial denominator or independent fleet-reliability measurement established. ASIMOV-Jailbreak evaluates adversarial prompts; Argon 99.5% result is not physical accident probability. Avoid 30–100% headline uplift without heterogeneous task baselines.
- NVIDIA — CoRL robotics announcements
November 9 GR00T 2 announcement and partner examples. The linked Isaac Lab 3.0 release is dated October 20; its software release is not redated as a new November 9 event.
- NVIDIA Research — GR00T 2 evaluation
NVIDIA reports average trial success across 50 tasks of 40% zero-shot, 71% with one demonstration per task and 95% with five. Succeeding at least once on 40 of 50 tasks is a different denominator. Code is marked for December; no independent replication.
- Isaac Lab — actual 3.0.0 release
The actual v3.0.0 release page is dated October 20. Available background tooling, not a new November 9 launch.
Images & figures
GR00T 2 average trial success across 50 tasks rises from 40% with no task demonstration to 71% with one and 95% with five.
Euangelion · Source · Usage terms