Robot VLA claims now live or die on execution budgets

VLA papers treat action generation as a control design problem. Libra-VLA separates coarse discrete intent from continuous fine control, then runs the planner less often while the refiner acts at the control rate. It reports 97.2% average success on LIBERO and 79.5% on LIBERO-Plus zero-shot transfer.

World-action work adds future-state prediction to the same constraint. MotuBrain predicts actions and future visual states in one diffusion model, with multiview support and a faster inference stack. Its reported latency falls from 4.90 seconds to 0.09 seconds, raising frequency to 11.11 Hz. The useful claim is practical: predictive models matter when they fit the robot’s update budget.

Several papers add decision checks during rollout. VLA-ATTC spends extra compute only when action samples disagree, then uses a Relative Action Critic to pick among candidates. On real Agilex Piper tasks, PI0 rises from 46.0% to 58.7% success with the adaptive version, while the paper reports 20.8 Hz control.

Sentinel-VLA adds a status monitor that detects Initial, Normal, New-subtask, and Error states. It plans at the start, reuses memory during normal execution, and generates recovery behavior after errors. The reported 13 ms/action latency keeps the monitor close to PI0 timing while improving real-world Agilex Piper success to 60.0% average across three tasks.

Navigation work shows the same concern under network delay. AsyncShield realigns delayed cloud VLA waypoints with an SE(2) pose transform and adds a safety-constrained local policy. Under mixed network degradation, success reaches 76.7%, while the no-alignment ablation drops to 36.7%.

The week’s deployment papers treat released robots as data sources. LWD trains a single VLA policy with offline data, autonomous fleet rollouts, and optional human interventions. Its real-world evaluation uses 16 dual-arm robots across 8 tasks, including 3–5 minute long-horizon manipulation, and reports 95% average success after a few hours of online interaction.

Data collection work attacks the smaller-lab bottleneck. Phone2Act turns an Android phone into a 6-DoF teleoperator and records synchronized demonstrations in LeRobot format. Fine-tuning GR00T-N1.5-3B on 130 collected episodes gives 9 successes in 10 real Dobot CR5 trials, though the phone-to-robot path still runs at a measured 350–440 ms latency.

Lucid-XR broadens the data path with headset-based simulation and generated multi-view images. In 30-minute sessions, users collected about twice as many demonstrations as real teleoperation, and augmentation raised effective dataset size to about five times the real baseline.

Simulation papers focus on contact, visual fidelity, and hardware transfer. GS-Playground combines a batched 3D Gaussian Splatting renderer with a parallel physics engine, then binds rendered Gaussian clusters to rigid bodies during motion and contact. It reports about 10,000 FPS rendering at 640×480 and support for up to 2048 rendered scenes at that resolution.

DexSim2Real uses a vision-language model as a realism critic to tune simulation randomization for dexterous manipulation. The final policy trains in Isaac Sim and transfers to a real Franka Panda with an Allegro Hand. Across six real tasks, it reports 78.2% average success and an 8.3% sim-to-real gap, with ablations showing losses when tactile input, guided randomization, or the skill curriculum are removed.