What Google DeepMind announced with Gemini Robotics 2
On 30 July 2026, Google DeepMind announced Gemini Robotics 2, a suite of three physical AI models built to act as a universal intelligence layer for humanoid and autonomous robotics. The release moves beyond earlier upper-body manipulation systems by introducing whole-body control under a single learned policy, meaning a humanoid robot can walk, crouch, balance and perform fine manipulation tasks simultaneously rather than through separate, stitched-together control systems. DeepMind describes this as control “from feet to fingertips”.
For environmental consultants, developers, councils and legal teams, this matters less because of the humanoid demonstration itself and more because of the strategic model DeepMind is pursuing. Rather than building robots, DeepMind is positioning its software as the intelligence layer that third-party hardware manufacturers, such as Apptronik and Franka, can install into their own machines. That is a meaningful shift for anyone thinking about how autonomous or semi-autonomous equipment might eventually be procured, insured and regulated on Australian sites, including contaminated land, remediation and construction environments where robotics vendors are increasingly pitching site inspection and monitoring tools.
The announcement was demonstrated primarily on Apptronik’s Apollo 2 humanoid and Franka’s F3 Duo bi-arm robotic system. DeepMind also released an updated safety benchmark, ASIMOV-Agentic, alongside the reasoning model, which is a signal that physical collision risk and human-adjacent operation are being treated as first-order engineering problems rather than afterthoughts.
The three Gemini Robotics 2 models, performance figures and safety benchmark
Gemini Robotics 2 comprises three distinct models working together. The first, Gemini Robotics 2 (VLA), is a Vision-Language-Action model that converts multimodal inputs, including images, video and voice, directly into motor control commands for full humanoid bodies and multi-fingered hands. The second, Gemini Robotics ER 2 (Embodied Reasoning), handles high-level task orchestration across hundreds of sub-steps spanning several minutes, with real-time video progress tracking and the ability to query external cloud tools such as Google Search when an instruction needs clarification. The third, Gemini Robotics On-Device 2, is a lightweight edge model that runs locally on the robot’s own hardware and can reportedly adapt to an entirely new robot form factor within hours using fewer than 200 demonstration examples.
DeepMind has published specific performance figures for these models. The system achieved a 76.3 percent success rate on shelf object retrieval tasks and 91.3 percent precision in frame-level video moment detection, which is the model’s ability to identify the exact point in a video where a specific action or event occurs. This sub-second moment finding underpins one of the more operationally significant claims in the release: mid-execution error recovery. If a robot drops an object, encounters a spill, or hits an unexpected obstacle partway through a multi-step task, ER 2 is designed to detect the failure and resume from the last valid step rather than resetting the entire workflow from the start.
The suite also introduces heterogeneous multi-robot collaboration, allowing machines of different physical form, for example a legged humanoid and a wheeled rover, to operate within a shared semantic space. In practice this means diverse hardware can hand off tools, divide multi-stage tasks and coordinate movement through the same physical workspace without each unit running incompatible proprietary control software.
Safety governance is addressed through the ASIMOV-Agentic benchmark, released alongside ER 2, which is used to evaluate physical collision risk, load limits and safe operation in spaces shared with humans. DeepMind has not published this as a certification standard but as an internal evaluation framework, and it is worth noting that no independent regulator, in Australia or elsewhere, has yet adopted or referenced it as a compliance benchmark.
Australian context
There is no direct Australian regulatory framework governing physical AI or humanoid robotics at this stage, and this release does not trigger any change to existing law. The relevant frame for Australian businesses is the existing Work Health and Safety Act and Regulations in each state, which impose general duties on persons conducting a business or undertaking to eliminate or minimise risks to health and safety so far as is reasonably practicable. Any deployment of autonomous or semi-autonomous physical AI on an Australian worksite, whether industrial, construction or environmental, would need to be assessed against those existing general duties rather than any robotics-specific standard, because none currently exists.
Hardware agnosticism is the aspect most relevant to Australian procurement decisions. Because DeepMind is positioning itself as a software layer rather than a hardware manufacturer, Australian operators considering autonomous equipment for logistics, industrial monitoring or site inspection work would not necessarily be locked into a single vendor’s proprietary control stack. This has implications for how contracts, warranties and liability are structured when equipment from different manufacturers is expected to interoperate under a shared intelligence layer, a point that legal teams advising on procurement or lease arrangements involving autonomous equipment should start turning their minds to now, well ahead of any Australian-specific standard being developed.
The reduced training data threshold, fewer than 200 demonstration examples for adapting to a new robot form factor, also lowers the barrier for smaller Australian operators or research institutions to trial custom robotic applications, rather than requiring the scale of data collection previously associated with large industrial robotics programs. For environmental work, that could make purpose-built robotic tools, such as repeat inspection of remediation sites, monitoring of capped landfills or assessment of areas unsafe for human entry, accessible to mid-sized consultancies and councils rather than only major contractors. Any such deployment would still need to be assessed against existing WHS duties, and operators should watch for the eventual emergence of robotics-specific standards or regulator guidance before committing to large-scale rollouts.
References and related sources
- Primary source: deepmind.google
- siliconangle.com
- robozaps.com
- blog.google
- automate.org
How iEnvi can help
iEnvi integrates technology and data-driven approaches into environmental consulting. We monitor AI and technology developments that affect how environmental professionals deliver services to clients.
This is an iEnvi Machete news summary. Prepared by iEnvi to summarise the source article for environmental professionals tracking AI, data, and technology developments that affect consulting and project delivery.
Published: 01 Aug 2026
Need advice on this topic? Speak to an iEnvi expert at info@ienvi.com.au or 1300 043 684, or contact us online.
Need advice on this issue? iEnvi provides practical, senior-led environmental consulting across contaminated land, remediation, ecology and environmental risk.
Contaminated land services Remediation services Talk to iEnvi