VLA deployment work targets latency, continuity, and scarce interaction data together

Three papers attack different sources of control delay. Temporal-redundancy removal caches stable visual tokens and compresses flow sampling to two steps, reaching 8.2 FPS on LIBERO with 93.8% mean success versus 94.4% for the original policy. Jetson-PI instead predicts a future representation to correct stale asynchronous observations; system optimizations raise Jetson Orin control from 0.70 Hz to 6.06 Hz. ChunkFlow complements raw speed with seam-aware training and overlap blending, reporting 93.4% on LIBERO-Long and 4.43 ms reasoning latency. Together, these results show that practical VLA control depends on coordinating perception reuse, action generation, hardware scheduling, and chunk execution.

The scarce-data signal continues, now with mechanisms that manufacture diversity rather than merely adding demonstrations. WANDA converts one RGBD demonstration into trajectories across reconstructed and generated 3D scenes; in simulation it reaches 75.6% average success, near a baseline trained on roughly 40–60 demonstrations. ExToken uses behavioral clusters to diversify reinforcement-learning rollouts: 256 rollouts reach 93.4% success, compared with 90.3% for the matched baseline and performance comparable to its 512-rollout setting. FlowWAM adds a complementary route: optical flow extracted from action-unlabeled video improves RoboTwin Clean success from 82.40% to 92.94%.

Explicit representation design remains a strong companion to efficiency. VistaVLA grounds semantic features in 3D Gaussian primitives, then compresses about 100,000 primitives into 64 policy-facing tokens. It reports a 22.8-point average real-world success gain across seven tasks, although large position shifts remain difficult. FlowWAM represents actions as optical-flow videos, allowing one pretrained video architecture to support both control and future prediction; it reaches 92.94% and 92.14% success on RoboTwin Clean and Random. The common result is narrower than a general move toward 3D or video: representations help when their coordinates and temporal structure match the control problem.