Vision-language-action papers focus on deployable robot control

Several papers make VLA policies easier to run as long-horizon controllers. PrimitiveVLA cuts demonstrations into 11 reusable primitives and plans primitive sequences at test time; it reports OpenVLA success rising to 45.50% on LIBERO-90-Novel and pi0.5 success reaching 80.25% on LIBERO-Long. ProgVLA uses a 0.1B-parameter policy with compressed multimodal tokens and progress heads, reaching 91.1% average LIBERO success and 68% real-world success over 100 PiPER-arm trials. A separate probing study finds that frozen VLA features already contain a value-like success signal, then uses a linear probe to select among Pi0.5 action chunks, improving pooled hard-3 success to 42.44% against 31.11% for greedy decoding.

VERA keeps planning in video space and trains a robot-specific Jacobian inverse dynamics model to convert generated frame-to-frame motion into actions. The method reports closed-loop simulation success of 70.0% on Allegro-Sim, 94.0% on Panda-Sim, and 92.5% on PushT-Sim, with the same video planner paired to different robot adapters. Mag-VLA applies a related execution idea at microscale: it fine-tunes Qwen2.5-VL-7B for bimanual magnetic microrobot control, predicts task phase, and emits short dual-arm action chunks. In real tests, approach success is 90% across three tasks, while transport success drops with harder paths.

The tactile papers treat contact quality as part of the control problem. Tabero builds vision-touch-language data by replaying manipulation trajectories in Isaac Lab with tactile sensing, then predicts pose and force targets for a hybrid controller. It reports over 70% lower average grip force under gentle instructions, while tracking peak and average grip and applied forces. CoP uses a center-of-pressure contact description for dexterous sim-to-real learning on an Allegro hand. On real peg-in-hole insertion, CoP reaches 0.78 success across six shapes, ahead of raw taxels at 0.48 and binary contact at 0.53.

Ω-QVLA compresses both the language backbone and diffusion action head of VLA policies to uniform W4A4 precision without training. On Pi 0.5 it reports 98.0% average manipulation success, close to the FP16 reference at 97.1%, while reducing static memory footprint by 71.3%. ReasonBreak covers a different deployment risk: small text corruptions in driving commands can alter reasoning and trajectories in NVIDIA Alpamayo models. In closed-loop simulation, the paper reports up to 72% attack success for trajectory manipulation and links successful attacks to higher collision, off-road, and wrong-lane failures.