Xiaomi-Robotics-1 reached a 75 percent average task-success rate with less than 10 hours of task-specific training after pretraining on roughly 100,000 hours of human-collected motion data.
Here's what's really happening
1. The Decoder reports Xiaomi trained its robot policy on handheld-gripper footage rather than robot rollouts, and its tests found that adding data improved results more than increasing model size. 2. The Decoder and TechCrunch report that Google is developing Frozen v2, a server chip that would hard-code parts of Gemini's architecture into silicon. Sources cited by The Information project 6-10x greater inference efficiency than Google's current TPUs, with deployment planned for 2028. 3. Microsoft plans to use AMD's Helios platform in Azure after it ships in the second half of 2026. Separately, a public GitHub profile belonging to an AMD director suggests Anthropic is testing AMD hardware, according to The Decoder. 4. Hugging Face published Grabette, an open, low-cost handheld system for recording robot-manipulation demonstrations and converting them into robot-ready datasets.
Builder/Engineer Lens
Frozen v2 reportedly embeds parts of Gemini's architecture rather than its trained weights, so new weights could still be loaded while the chip remains tied to a narrower model structure. That trade exchanges flexibility for lower serving cost and latency if Google keeps the architecture stable. Xiaomi and Grabette point to a parallel specialization on the data side: teams can collect manipulation demonstrations with instrumented handheld grippers instead of using a physical robot for every recording. Helios adds another potential cluster option, but it is not yet generally available; the practical work now is keeping training software portable enough to benchmark it when Azure exposes the hardware. Together, these developments shift attention from parameter count alone toward data throughput, workload portability, and hardware-software co-design.
What to try or watch next
- Pilot a Grabette-style collection workflow and measure task success against additional hours of demonstrations before increasing model size. - Establish latency, power, and cost baselines for current inference workloads so future specialized hardware can be evaluated against evidence rather than headline efficiency projections. - Keep training jobs portable across accelerator backends and benchmark Helios when Azure makes instances available; do not assume the announced platform is deployable today.
The takeaway
Today's signal is not that scale has stopped mattering. It is that AI teams are finding another competitive path: specialize how data is collected and how models are served, then measure whether the efficiency gain justifies the loss of flexibility.