Stable Bodies, Rolling Models: The Robot Platform Split

Two robotics releases this month look like they belong to different industries, but they describe the same product shift. Boston Dynamics rebuilt its trailer-unloading Stretch robot around what broke in the field; Google DeepMind put a vision-language-action model on an active Mercedes-Benz production line. In both cases the physical machine is being engineered for stability while the intelligence is delivered as a rolling software upgrade.

The dock robot gets a reliability budget

Stretch Gen 2, unveiled November 3, is explicitly a field-data design. Boston Dynamics says its team mined more than 100,000 operating hours and tens of millions of unloaded boxes to target the most common causes of unplanned downtime S1. The headline change is subtractive: cutting 81 percent of internal cabling — about half a mile of wire — because intermittent electrical faults and wiring wear were the biggest reliability tax. Heavy-duty automotive-grade connectors handle vibration, continuous-rotation joints eliminate hard stops that accumulated mechanical stress, and exterior panels are less likely to trip a false safety stop when a box falls.

Service speed got the same treatment. A full swerve drive swap now takes about an hour instead of a full day, onboard diagnostics push routine maintenance to the customer's own team, and throughput stays at up to 1,000 cases per hour with a battery rated for two shifts S1. Just as notable is what does not change: case detection, motion planning and pick strategies improve across the whole fleet through regular software releases, with no hardware modifications required.

The model layer graduates to a live line

On November 10, Google DeepMind introduced Gemini Robotics 2.5 and reported a 96 percent success rate when Apptronik's Apollo 2 humanoid picked and packed automotive parts on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany S2. The architecture splits the job: the Gemini Robotics ER embodied reasoning model keeps track of the workflow, while the 2.5 actions model controls the hands and arms reaching into supply bins, picking metal brackets, and placing them into sequencing containers, correcting for gaps S2. Across an evaluation suite of more than 200 distinct tasks, DeepMind reports 30 to 100 percent higher success rates than previous baselines and 20 to 30 percent faster completion; the release also introduces ASIMOV-Jailbreak, an open-source benchmark of disguised unsafe prompts for testing resistance to adversarial jailbreaks S2.

The family also includes On-Device 2, a lightweight VLA "optimized to run locally on robotic hardware," which matters for cells that cannot depend on a network round-trip S3. Google's announcement also relays Apptronik's report that only a few weeks of post-training produced a substantial performance gain, with almost no drop-off when deployed on the Mercedes-Benz site S4.

What buyers and builders should ask

This is analysis, not reporting: the two announcements together suggest procurement questions are changing. For hardware, ask what a swap costs in hours, how much wiring remains, and whether diagnostics are deep enough for in-house crews — that is where uptime is won. For models, ask whether claimed numbers come from lab benchmarks or active production lines, and whether the control loop runs on-robot or in the cloud. A 96 percent proof-of-concept is not a deployment contract, and kitting is a narrow task even across two plants. But the direction is clear: the body is bought once and serviced, while capability now compounds through model updates. Evaluate robots the way you evaluate software platforms — by upgrade cadence as much as by spec sheet.