VLA work now judges policies by grounding, memory, and action checks

Several robot papers add intermediate target signals between language and motor control. AVP makes the vision-language model output visual primitives before action prediction, then feeds those tokens to a flow-matching action expert. On Chinese chess manipulation, it reports 90.28% average success versus 62.67% for π₀.₅, with 0.27 seconds per instruction.

GesVLA handles a different source of ambiguity: human pointing. It turns wrist and index-finger keypoints into gesture tokens and combines them with language and scene perception. Across three real-robot tasks, success reaches 83.3%, compared with 31.7% for a text-only VLA. SOMA adds persistent spatial memory for objects outside the current camera view, using head-camera scans and memory tokens with semantic and 3D position data. Its gains are smaller in success rate, but it cuts target search time and grasp attempts across out-of-vision tasks.

A second group works on the state that policies use after the first observation. EvoScene-VLA carries a recurrent scene prefix across control calls and predicts future scene tokens with the action chunk. On 31 RoboTwin tasks, it raises average success to 89.1% in fixed scenes and 88.5% with randomized layouts.

LVDrive applies a similar idea to camera-only autonomous driving. It predicts future scene features in latent space and uses them to refine the trajectory, reaching 80.71 Driving Score and 58.26% Success Rate on Bench2Drive. The TRM paper shows a related planning issue: a latent world model may contain the right state, while the terminal cost ranks candidates poorly. Replacing raw latent distance with a learned reachability metric raises hard TwoRoom success for LeWM from 7.0% to 97.0%.

CrossVLA treats VLA adaptation as an empirical systems problem. It adapts Direct Preference Optimization (DPO) to both autoregressive action-token policies and flow-matching continuous-action policies. On OpenVLA across LIBERO’s four suites, DoRA+DPO reaches 73.2% mean success versus 62.75% for supervised fine-tuning. Its latency breakdown also matters: π₀.₅ spends about 78.6% of sample-actions time in the denoising loop, so prefix key-value caching is a weak target.

Pre-VLA adds a runtime checker before execution or world-model rollout. It predicts action validity and critic-derived advantage for candidate action chunks. On LIBERO, it reports 0.9542 validity accuracy and raises RynnVLA-002 average closed-loop success from 30.79% to 37.62%, with 183.9 ms verification per chunk.

The endoscopic surgery copilot paper extends VLA reasoning into a high-risk setting, but its evidence is architectural and clinical rather than experimental. It defines a Level of Autonomy 2–3 assistant that can generate options, monitor the surgical scene, and execute bounded low-level maneuvers under surgeon authority. The proposed system fuses endoscopic video with preoperative and intraoperative signals, including CT/MRI priors, ultrasound, OCT, tracking, and force proxies.

The paper is useful as a requirements document. It names specific subtasks such as traction, dissection, hemostasis, smoke response, and bleeding-point localization. It also sets a sub-second response target. It reports no benchmark, runtime measurement, clinical trial, or success rate, so it should be read as a safety specification for future VLA work rather than a deployed result.