Longer robot memory and faster prediction become practical control mechanisms

RoboTTT compresses up to 8K visuomotor timesteps into fast weights, keeping latency constant as history grows. It reached 79% average completion on three real-robot assembly tasks, versus 42% for its single-step baseline. FoMoVLA takes a complementary route: it trains compact future-state tokens together with sparse point trajectories, then removes most auxiliary machinery at inference. It reached 98.8% average success on LIBERO with 9.4 ms median overhead. Together, these studies make temporal information operational through compact internal state rather than an ever-growing observation buffer.

DriftWorld generates action-conditioned futures in one forward pass at more than 30 frames per second, averaging a 17× speedup over diffusion baselines across five benchmarks. Reflex similarly avoids repeated work inside flow-matching vision-language-action (VLA) policies. Its partitioned cache and asynchronous pipeline cut Pi0.5 inference on LIBERO from 135.2 ms to 52.4 ms and support stable 50 Hz streaming. The shared result is narrower than general real-time autonomy: iterative generative components can be reorganized so planning and execution no longer wait on full recomputation.

Aggregate success can conceal brittle behavior. FLARE reduced baseline success to zero on three LIBERO suites using optimized physical lighting. It also found that broad color augmentation could appear robust by teaching policies to ignore color; its chroma-preserving defense restored 97.5% benign and 92.5% attacked success on one real color-dependent task. A separate active-testing framework sampled structured variations in pose, table height, and viewpoint. Across 2,331 real-world evaluations, it typically reduced the trials needed to characterize performance by 20–40% compared with random testing. Both papers argue for testing failure regions and retained capabilities, not only average task completion.

Two sources use generated consequences to improve a policy without making the generator the deployed controller. Xiaomi reports that adding scenes and videos from its open 38-billion-parameter U0 world model raised π0.5 success on unfamiliar manipulation from 36.9% to 63.2%. In contact-rich manipulation, a Latent Tactile Predictor trained intermediate action features to anticipate future touch, then was removed for inference; average real-world success reached 74%. The mechanisms differ, but both place outcome modeling in the learning pipeline. The Xiaomi gain is vendor-reported, while the tactile result covers five tasks and two VLA backbones.