Most lab robots follow rigid scripts. They repeat the same motion, in the same order, and stall the moment something shifts on the bench. Google’s Gemini Robotics team is trying to change that. In a technical report posted to arXiv, the team describes Gemini Robotics-ER 1.5, an updated model designed to give robots the ability to survey a cluttered workspace, build a plan, execute physical tasks, and track whether each step actually worked before moving on. The update, detailed in spring 2026 documentation, targets a persistent bottleneck in laboratory automation: getting machines to reason through problems instead of just executing memorized routines.
A two-layer brain for robots
The system splits decision-making into two tiers. On top sits Gemini Robotics-ER 1.5, the reasoning layer. It handles visual perception, spatial mapping, task sequencing, and progress estimation. Below it, a vision-language-action (VLA) controller converts those high-level decisions into precise joint movements and gripper commands. Think of it as a division between the strategist and the hands: one figures out what needs to happen next, the other makes it happen physically.
This architecture builds on a foundational report the same team published earlier, which introduced capabilities like multi-view correspondence and 3D bounding boxes. Those features let a robot stitch together camera feeds from different angles to build a spatial model of its surroundings and pinpoint where objects sit in three-dimensional space. Without that foundation, even basic tasks like picking up a flask on a crowded bench or navigating around an obstacle become unreliable.
The ER 1.5 update pushes further into what the authors call “embodied reasoning tasks,” a category that bundles spatial understanding with planning and real-time self-assessment. In practice, that means a robot running the model can look at a lab setup, generate a sequence of steps, and continuously verify whether each step succeeded. If a pipette slips or a container is not where it was expected, the system can recognize the failure and adjust rather than freezing or blindly continuing.
What the reports actually show
Both documents were published on arXiv, the open-access preprint repository run by Cornell University. That choice matters for two reasons. First, it makes the full technical details available to competing labs, hardware makers, and independent reviewers without a paywall. Second, it means neither report has gone through formal peer review, so the methods and results have not yet been vetted by outside experts through a journal’s editorial process.
The reports describe model architecture, training methodology, and benchmark evaluations chosen by the authors. They do not include data from live laboratory deployments. There are no published success rates from active biotech, chemistry, or physics labs, and no documentation of how the system handles the kinds of surprises that define real research environments: unlabeled reagents, equipment that has drifted out of calibration, or a colleague who rearranged the bench overnight.
No Google executives or DeepMind representatives have publicly commented on deployment timelines, partner institutions, or pricing. The reports are technical documents, not product announcements, and they stop short of describing any commercial rollout.
Where it fits in the robotics race
Google is not working in isolation. Companies like Figure AI, Tesla (with its Optimus humanoid), and 1X Technologies are all pursuing general-purpose robots that can operate in unstructured environments. What distinguishes the Gemini Robotics approach is its emphasis on pairing a large multimodal model’s reasoning ability with a dedicated action controller, rather than training a single end-to-end system to handle both thinking and moving.
That architectural bet carries trade-offs. A two-layer system can, in theory, swap in improved reasoning models without retraining the entire action stack. But it also introduces a coordination challenge: the reasoning layer and the VLA controller need to stay tightly synchronized, and any latency or miscommunication between them could cause errors in time-sensitive tasks. The reports acknowledge this design but do not publish latency benchmarks or failure-mode analyses that would let outsiders evaluate the trade-off directly.
Why independent replication is the next milestone that matters
For research facilities weighing whether AI-driven robotic assistants are ready for real work, the ER 1.5 report signals that the underlying technology is advancing quickly. Spatial reasoning, task planning, and progress-aware control are no longer theoretical goals; they are implemented capabilities with documented architectures. But the gap between benchmark performance and reliable daily use in a working lab has not been publicly bridged.
The safety discussion in the foundational report acknowledges risks associated with autonomous decision-making in spaces shared with human researchers, but neither document references external ethical reviews, regulatory filings, or third-party audits. Whether internal oversight processes exist but remain unpublished, or are still being developed, is unclear from the available materials.
Until outside robotics groups test the system’s claims against their own benchmarks, or until Google publishes results from pilot deployments in partner labs, the Gemini Robotics-ER 1.5 system is best understood as a promising research platform rather than proven infrastructure. The technical vision is clear and ambitious. The evidence that it works outside controlled conditions has yet to arrive.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
