Robots learn in minutes, cars ship the tech, and safety still can't tell a collision

Seven minutes to a boxer

HoloBoxing, described in a preprint, learns humanoid boxing by combining a frozen whole-body controller with a compact planner that predicts body-point targets. Measured-state imitation initializes the planner before reinforcement learning refines interaction behavior, and the recipe accounts for about seven minutes of additional training. That figure excludes controller pretraining, data preparation, evaluation, and search, but the point stands: once the controller exists, teaching a new physical skill is a small optimization, not a new training run. m:…2139

RynnUnify, another preprint, pushes the same consolidation further: a 16B unified embodied model that integrates visual planning, physical reasoning, and robot control in a Mixture-of-Transformers architecture with an autoregressive Reasoner and a rectified-flow video Generator. Its world-action extension achieves 51.5% success on 20 selected composite RoboCasa365 tasks, and it achieves 46.7% average success across three real-robot tasks, with its agentic extension improving a reasoning benchmark average by 6.0 points over direct reasoning. m:…6660

Echo, a third preprint, addresses a quieter problem in robot control: knowing when a task has stopped, failed, or is still going. It represents task status with stop and failure signals added as two dimensions of each action chunk and denoised together with robot actions, learning these signals without fine-grained progress annotations. Real-world manipulation experiments show it accurately predicts task status and consistently improves policy performance, which matters because a robot that cannot tell it has failed will keep failing. m:…2950

The safety gap

The counterweight is TouchSafeBench, a preprint that tests whether vision-language models can ground collisions in human-robot collaboration scenarios. Current frontier and robotics-oriented models achieve less than 50% Macro-F1 on safety-critical collision classification and warning tasks across 2,940 simulated episodes, robot-scene contact is consistently harder for them to detect than human-contact risk, and explicit depth information is not automatically transformed into robot-body collision evidence. m:…6983

That is the dangerous combination in this week's embodied AI news: training a new skill is getting cheap, unified models are consolidating planning and control, and the models still cannot reliably tell when they are about to hit something. Visual fluency is not physical accountability, and deployment timelines that assume otherwise are borrowing against safety. m:…6983 m:…2139

Already in the driveway

Horizon Robotics says the Jetta M6 was officially launched on October 22, 2026, positioned as an ultra-comfortable AI pure electric family sedan priced between 79,800 and 99,800 yuan. The 555 Ultra version uses the Horizon Journey 6M chip and HSD AI foundation model with the Hyper Sense 8 one-stage end-to-end in-vehicle solution, supporting highway and urban navigation on autopilot plus full-scenario intelligent parking, without high-precision maps. m:…8354

The scale behind the launch is the real story: in 2026, CoreTech's full-scenario advanced driving assistance solution was equipped in 7 new electric models from Volkswagen's three joint ventures in China, the Horizon Journey series has achieved a mass production scale of 15 million units, and the technology is implemented across 40+ automotive brands and 350+ mass-produced models. HSD is described by Horizon as the first mass-produced one-stage end-to-end model in China. m:…8354

Embodied AI's mass-market moment is arriving through cars first, where the safety question is literal and the volumes are real. The research question from TouchSafeBench travels with it: the systems getting cheaper to train and faster to ship still need to prove they understand physical contact before they share space with people at scale. m:…8354 m:…6983