Closed-loop execution becomes the decisive test for embodied AI

IMBench makes the reasoning–action gap measurable. Leading vision-language models reached roughly 74% constraint understanding, but GPT-5.5 completed only 11.3% of tasks from vision and 18.8% with privileged object state. Several alignment, tool-use, hidden-state, and balancing tasks remained at zero.

AC-VLA addresses a related failure at the policy level. It decomposes instructions into reusable sub-tasks and masks wrist views during selected phases to reduce trajectory memorization and visual shortcuts. On LIBERO-OOD, the π₀.₅ variant reached 64.2% Spatial-OOD and 73.3% Goal-OOD success, gains of 28.7 and 26.7 percentage points. Together, the studies support evaluation and training around executable recombination rather than verbal or in-distribution competence alone.

Fast-slow driving turns latency into an architectural choice. A frozen 7B backbone refreshes scene context at 5 Hz, while a 337M action expert uses the latest frame to issue control at 20 Hz. CARLA route completion rose from 37.0% with replayed commands to 94.0%; however, the long-route driving score was only 2.96, leaving safety unresolved.

JoyNexus applies a similar separation principle to post-training infrastructure. It keeps a shared backbone resident while isolating tenant-specific action modules, optimizer state, and policy versions. Group batching reuses backbone computation across compatible workloads. The paper reports lower aggregate GPU time but gives no numerical efficiency values, so this is an architectural signal rather than a quantified systems advance.