Robot VLA progress is measured by control-loop success

Generalist robot policies are being trained to share high-level manipulation concepts while still producing robot-specific controls. ZR-0 is the clearest example. It uses dense embodied chain-of-thought labels during training for scene description, task progress, future plan, target boxes, and discrete action tokens. At inference, it skips text generation and outputs continuous action chunks through a diffusion action expert.

The measured claim is concrete. ZR-0 reports 97.8% average success on LIBERO, with ProcCorpus-60M covering about 60 million frames, 1,000 hours, and more than 400,000 trajectories. The same daily trend also groups this with reward-free test-time improvement and trajectory memory, showing that policy scale is being tied to executable action timing and state history.

Several papers add explicit physical signals before action generation. 3D HAMSTER predicts end-effector waypoints in metric 3D and feeds them to a point-cloud control policy. On DroidSpatial-Bench, it reaches 65.5% Both accuracy at 10 cm, above RoboBrain-2.5-8B at 60.1% and far above general VLM baselines.

Bridge-WA uses a different control signal. It distills a future-change teacher into intended-outcome tokens, change maps, and motion-flow maps, then removes the teacher at deployment. It reports 52.8% average success on VLABench and stronger real-robot Dobot hard-track averages than X-VLA under distractors, lighting changes, and tablecloth changes.

DVG-WM adds video prediction evidence. It separates low-resolution dynamics from high-resolution refinement and reports 88.7 seconds inference on LIBERO video prediction, compared with 236.8 seconds for CogVideoX-5B and 354.2 seconds for LVP-14B. The common pattern is direct: useful robot prediction is judged by contact, changed regions, and planning cost.

Long tasks are exposing failure modes that short manipulation benchmarks miss. FurnitureVLA targets real-scale bimanual furniture assembly with 4 to 7 subtasks and up to 1,550 control steps. Its policy predicts both 14-dimensional bimanual actions and a continuous progress value, using the progress signal to trigger subtask transitions.

The payoff is measured on full assemblies. Average simulated success rises from 0.48 for a monolithic finetuned VLA to 0.80 across three IKEA-style tasks. The same result also shows the cost of weaker execution design: removing the rear-camera setup drops average success to 0.50, and discrete progress prediction fails all three simulated furniture tasks.

Robot foundation models are also being evaluated as deployed services. ROSA routes many robots’ model requests to a shared GPU pool and schedules them by service-level objective qualified action throughput. The task file covers model components, prompts, call rates, retry rules, and fallback actions such as stop, resend, replan, or call a human.

The reported gains come from treating robot inference as a fleet workload. On 8 NVIDIA H200 GPUs with up to 64 virtual robots, ROSA improves SLO-qualified factory productivity by up to 12.06x over dedicated serving baselines. Against shared-server baselines without its scheduler, it improves qualified factory action throughput by up to 2.44x. This makes deployment evidence part of the research signal, alongside policy success rates.