Robot VLA Work Concentrates on Generalization Under Real Control Conditions

Qwen-RobotManip treats robot diversity as an alignment problem. It maps different arms, grippers, cameras, and action spaces into a shared state-action template, then trains on about 38,100 hours of manipulation data. Most of that scale comes from a human-to-robot synthesis pipeline that renders egocentric human demonstrations into 15 bimanual robot configurations. The report claims first place on RoboChallenge Table30-v1 generalist track and reports real-robot validation across AgileX ALOHA, Franka, UR, and ARX platforms.

MuseVLA adds a different kind of generalization pressure: the policy must decide when RGB is insufficient. It selects thermal, acoustic, mmWave, or no extra sensor from the instruction and scene, converts the chosen measurement into a grounded sensor image, and feeds it back into the VLA. On real dexterous-hand tasks, synthesized pretraining reaches 80.6% average success on seen sensor-guided tasks and 66.7% on unseen tasks.

Several papers add internal prediction steps before robot actions are decoded. PearlVLA keeps action generation fast by refining latent plan tokens with a frozen latent world model, then decoding a continuous action chunk. It reports 98.7% average success on LIBERO, above the strong baselines listed in the local summary.

ThinkingVLA makes the intermediate reasoning visible to the model: text subgoal, predicted future image, action-focused rationale, then action. It reaches 77.9% average success on RoboTwin 2.0 Easy and reports 85% average success across five real ALOHA tasks. WAM-RL tests the training side of the same idea by updating both a video world model and an actor during online interaction. On LIBERO-Object, success rises from 68% for the base model to 82%.

Deployment-oriented papers focus on what a policy should do when its first action sample may be wrong or when the current image omits task history. VERITAS samples several short action chunks, scores them with a visual verifier, executes the best candidate, and later fine-tunes on successful verified rollouts. Across three policies and 1,160 evaluation episodes, it reports 12.6% average success gains in simulation and 35% gains in real-world deployment without policy fine-tuning.

Uncertainty Quantification for Flow-Based VLAs adds a confidence signal through velocity-field disagreement across action-head ensembles. The same signal drives SAVE, an active fine-tuning method that chooses uncertain cases for expert demonstrations. WeaveLA addresses repetition: it writes compact latent memory only at sub-goal completion events. On RoboMME SwingXtimes with N=3, success rises from 0% to 47.8% with a pi_0.5 backbone in the reported setting.

Evaluation work stresses detailed diagnosis. EBench tests 26 mobile, long-horizon, and dexterous simulation tasks across five capability axes and four generalization shifts. The headline result is that π0, π0.5, XVLA, and InternVLA-A1 sit in a narrow overall test-success band, 24.4% to 29.5%, while their skill profiles diverge sharply. InternVLA-A1 performs well on mobile manipulation but drops to 5.8% success on dexterous fixed-base tasks.

WireCraft targets industrial deformable linear objects such as wires and cables. Privileged state RL solves several simulation settings, including 95.86% insert success for Ethernet connector insertion with State PPO. Vision-based methods remain much weaker: Vision PPO reaches 17.74% insert success on the same Ethernet task, and simulation-only ACT gives 0/10 real UR5 insertions in the reported test.