Robot VLA papers are converging on rollout-grounded reliability

ABC raises the reproducibility bar for behavior cloning by releasing data, code, hardware details, model weights, simulation assets, and real rollout scores. Its ABC-130K dataset contains 3,553 hours, 134,806 episodes, and 195 tasks, with more than 100 hours of real policy rollouts in ABC-Eval. The paper also reports that simulation and some offline diagnostics track real-world success, giving researchers cheaper signals for model choices.

LA4VLA attacks a different data problem inside VLA pretraining. Long demonstrations often pair one high-level instruction with many image-action frames, so the language signal is sparse. LA4VLA cuts demonstrations into short language-action segments, keeps 33,116 human-verified episodes, and trains a 1B model with image-free language, proprioception, and action trajectories before VLA training. Its mixed language-action and VLA pretraining reports 83.3% real-world manipulation success and a 45.0 point gain over no pretraining on real tasks.

Several papers treat a frozen or trained VLA policy as one component inside a monitored execution loop. E-TTS samples reasoning-action pairs, scores them with vision-language verifiers, stores recent history, and asks for feedback when candidates fail. Across its reported settings, it gives an average simulated success gain of 13.52 points and a maximum gain of 33.14 points.

PhysReflect-VLA adds physical-feasibility scoring before action execution and compares predicted next states with observed next states after execution. If the mismatch is large, a reflector generates corrective guidance. On five real-robot tasks, Phys-OVLA improves average success by 5.4 points over OVLA-FT, and Phys-OFT improves by 3.0 points over OVLA-OFT.

RouterVLA shows that pre-deployment smoke tests can also guide policy choice. On LIBERO-Plus, a simple probe-success rule selects among frozen experts and reaches 0.6149 held-out success, compared with 0.4686 for the global best expert.

PAMAE modifies flow-matching VLA policies with phase-aware mixture-of-experts action heads. The router uses execution cues such as gripper state, gripper change, previous action norm, and progress. On five simulated multi-stage tasks, PAMAE improves π0 average success by 9.2 points and π0.5 by 5.6 points.

VibeAct adds a compact tactile channel for dexterous control. Piezoelectric fingertip microphones feed a contact and slip estimator, which supplies a 12-D tactile vector to policies trained in MuJoCo. On hardware, it improves success over a proprioception plus point-cloud baseline on Box Climb, Can Climb, and Nut Rotation.

The LeHome garment-folding system gives a competition-scale example for deformable objects. It combines a flow-matching VLA policy with reinforcement learning, replay, human corrections, and sim-to-real tuning, placing first in the online simulation round and second in the real-world final.

ForesightSafety-VLA makes safety the main measurement target for VLA policies. It defines 13 safety categories across physical interaction, instruction handling, and perception. The benchmark builds 66 safety-augmented RoboTwin scenarios across five robot embodiments, then separates safe success, unsafe success, safe failure, and unsafe failure.

The reported baselines all show remaining risk. OpenVLA-oft has the best listed safety-adjusted success rate at 0.35, while unsafe success still appears at 0.06. Among successful episodes, the unsafe share is 12.5% for OpenVLA-oft and 37.5% for ACT. That scoring choice matters because a robot can complete an instruction while colliding with objects, entering hot zones, or violating clearance limits.