One model acts, another thinks

Several robot and voice systems released between July and October 2026 share one shape: a fast model that acts or speaks, directed by a slower model that plans. Google DeepMind’s Gemini Robotics 2, OpenAI’s GPT-Live and Google’s Gemini 3.8 Live Extended Thinking all work this way. Robot policies also began to be served like other models, through a commercial API and, in research, from shared GPU pools. Evaluations were less flattering. A standard manipulation benchmark is close to saturated while its perturbed version is not, and in one safety report no tested model stopped reliably for nearby people without also stopping too often.

Scope: models, APIs and papers on vision, video, voice and robot control. Period: 1 July to 27 October 2026. Performance figures are the authors’ or vendors’ own unless stated otherwise.

16 July and 3 August: Xiaomi pre-trains a robot policy without robots

• What happened: Xiaomi described Xiaomi-Robotics-1, a vision-language-action (VLA) model pre-trained on more than 100,000 hours of manipulation trajectories recorded with UMI devices (arXiv). UMI is a hand-held gripper rig for collecting demonstrations without a robot (Chi et al., 2024). A vision-language model labels each clip with a description of the state change, so pre-training needs no manual labelling. Post-training on over 10,000 hours of cross-embodiment data then maps the skills onto real robots. On 3 August Xiaomi released code and checkpoints under Apache 2.0 (GitHub, Hugging Face).
• Why it matters: Xiaomi calls robot data the bottleneck for scaling policies. The paper reports steady gains with more pre-training data and larger models, and that stronger pre-training gave better out-of-the-box results on real robots. On RoboCasa365 it reports 57.4% success against a previous best of 46.6%.
• Confidence: high for the release. Scaling and benchmark results are Xiaomi’s own. The released code covers post-training, inference and evaluation, not pre-training.

30 July: Gemini Robotics 2 controls a whole humanoid

• What happened: Google DeepMind released three models (DeepMind). Gemini Robotics 2 is a VLA model that controls humanoids “from feet to fingertips”; the previous generation controlled the upper body for table-top tasks. Gemini Robotics ER 2 is a vision-language model that plans multi-step tasks, talks to people and directs the VLA. Gemini Robotics On-Device 2 runs on the robot and adapts to a new bi-arm body within a few hours, typically with fewer than 200 examples. ER 2 is in public preview in the Gemini API and supports blocking function calls for physical robot actions (Gemini API release notes). The VLA and on-device models are limited to early-access partners.
• Why it matters: the planner and the actor are now separate products, and only the planner is open to every developer. In DeepMind’s tests one checkpoint drove three robot bodies. DeepMind also states that multi-finger manipulation “remains challenging” and that movement speed needs work.
• Confidence: high for the release. Success rates come from DeepMind’s own evaluations.

29 and 30 July: a safety report that shows the trade-off

• What happened: with the release, DeepMind published a safety report and a benchmark, ASIMOV-Agentic (safety report, dataset). It tests whether the planning model refuses unsafe tasks, stops for nearby people, reacts to fault messages and asks for help when the VLA is likely to fail. Frontier models were reliable when they only had to classify a request as allowed or not in text. Results varied more when the same constraint had to become a point, a bounding box or a tool call.
• Why it matters: the human-proximity test puts numbers on a hard trade-off. Models that kept unnecessary stops under 5% missed more than 40% of real hazards. Models that missed only 10 to 15% stopped needlessly 15 to 25% of the time. The report recommends using these models “alongside deterministic, low-level safety guardrails” and states that it does not evaluate the functional safety architecture a real deployment needs.
• Confidence: high for what the report states. The benchmark is new and comes from the lab whose model it tests.

5 and 12 August: robustness, not LIBERO, separates VLA models

• What happened: LIBERO-Plus, published in October 2025, perturbs a standard manipulation benchmark along seven axes, among them camera viewpoint, lighting and the robot’s starting pose. Its authors saw success fall from 95% to below 30% under modest changes, and found that models tend to ignore the language instruction (Fei et al.). Two August papers report results on both versions. DreamWAM trains a world-action model to predict future motion, geometry and semantics as well as pixels. In one inference setting it lifts LIBERO from 97.3% to 98.4%, and LIBERO-Plus from 51.4% to 63.4%, over its pixel-only baseline (Yuan et al.). StellaVLA conditions the policy on one retrieved, structured demonstration and reports 98.8% on LIBERO and 85.1% on LIBERO-Plus (Xu et al.).
• Why it matters: LIBERO is close to its ceiling. In DreamWAM’s comparison the clean benchmark shows a gap of about one point, the perturbed one about twelve. When a VLA paper reports only clean-benchmark numbers, ask for the perturbed ones.
• Confidence: medium. Both are first-version preprints with evaluations run by the authors.

27 August and 1 September: video in both directions

• What happened: Google made Gemini Omni Flash generally available. It generates video from text and images, edits it through conversation and extends existing clips (Gemini API release notes). On 1 September Google added agentic video understanding. Instead of placing frames sampled once per second into context, the model reads the transcript first and loads only the visual segments it needs (Google).
• Why it matters: long video turns into a search problem, and the model pays for what it inspects. Google reports up to 88% fewer video input tokens and about 7% higher quality on long-form content. For clips under five minutes and latency-sensitive queries, it still recommends the static mode.
• Confidence: high for availability. The efficiency figures are Google’s.

10 and 15 September: voice models that talk while another model works

• What happened: OpenAI made GPT-Live 1 generally available. It is a full-duplex voice model: it listens and speaks at the same time and hands reasoning and tool use to a backend agent (OpenAI). Sessions cost $0.05 per minute, with backend model use billed separately (OpenAI changelog). Five days later Google released Gemini 3.8 Live Extended Thinking (Gemini API release notes), which plans and calls tools in the background while speaking short updates such as “Checking flight options now” (Google).
• Why it matters: this is the robot pattern applied to conversation, and it moves complexity into the application. OpenAI’s guide says the application “owns permissions, confirmations, business records, and task state” in either delegation mode (OpenAI). Google’s model can speak several times per request, so clients have to track an in-progress status rather than a single end-of-turn signal.
• Confidence: high. Both descriptions come from the vendors’ documentation.

1 July to 13 October: robot policies become a served workload

• What happened: two systems papers treat robot inference as a fleet problem. ROSA serves robot foundation models from a shared GPU pool over the network and reports up to 12.06x higher factory productivity than dedicated serving (Jiang et al.). Robion runs a VLA’s vision-language stage and action stage side by side on one GPU. On a 4-GPU server it serves up to 64 robots across 8 models while meeting latency targets for 98% of requests (Adamopoulos et al.). On 13 October Physical Intelligence announced a beta of its API, with access on request: customers upload data, fine-tune π models on managed compute and run them through hosted inference (Physical Intelligence).
• Why it matters: hosted inference puts the network inside the control loop. Physical Intelligence says most applications it showcased ran with cloud-hosted inference, and that some early deployments had network connectivity problems. Gemini Robotics On-Device 2 is built for the opposite case, robots that must work without a connection (DeepMind).
• Confidence: high for the API launch. The serving gains are the papers’ own measurements; ROSA’s are partly on synthetic workloads.

What to watch next

• CoRL 2026 runs from 9 to 12 November in Austin (CoRL). Speculation: more VLA papers there will report perturbed-benchmark results alongside clean ones.
• Access to the action models. Gemini Robotics 2’s VLA and on-device models are with early-access partners, and the Pi API is in beta (DeepMind, Physical Intelligence). The next step to watch is whether either opens to all developers.
• Safety tests under pressure. DeepMind plans to test whether models’ handling of safety messages holds with longer contexts and “attention jailbreaking” (safety report).
• Silent regressions in vision. On 25 September OpenAI fixed an image-encoding bug that had degraded image understanding in GPT-6 Sol and GPT-6 Luna, and advised customers to rerun their evaluations (OpenAI changelog).
• My read: the planner-actor split will become the standard interface for embodied and voice agents. That makes the hand-off between the two models the place to test hardest: what the planner asks for, what the actor can do, and who stops the robot when they disagree.