The context · July 2026
Google DeepMind’s July 2026 Gemini Robotics 2 announcement described work on whole-body robotic intelligence and an on-device model. [1]
A proposed project brief inspired by the period’s developments. This is an idea for exploration, not an announced Sansa product, an approved roadmap or a claim of completed client work.
Describe the states before labelling
Choose one handling sequence and define observable states such as approach, contact, lift and release. Add failed-attempt and uncertain states where appropriate. Avoid asking annotators to infer forces or intent that cannot be established from the available recording.
Capture the incomplete attempts
Include interrupted and unsuccessful sequences, not only attractive demonstrations. Have independent reviewers assess the boundaries between states. Keep examples from separate sessions in the evaluation set and document where camera placement limits what can be observed.
Deliver a useful account of behaviour
The proposed output is a labelled sequence dataset, a disagreement review and a list of observation gaps. It would support a controlled trial’s evaluation plan. It would not, on its own, certify the robot’s behaviour or establish that the system can operate safely outside the defined setting.
Source & context
Google DeepMind · Gemini Robotics 2, 30 July 2026This retrospective was written for the archive in September 2026. The linked primary source documents the announcement or event; the practical interpretation and proposed approach are Sansa’s editorial perspective. Public examples do not imply a client relationship. Product capabilities and guidance may have changed since the period discussed.
Another perspective · July 2026
Whole-body capability makes the task boundary more important
Continue readingWorking through a similar question?
Talk it through with Sansa