Robotics papers reward cleaner baselines and sharper tests of semantic control

StarVLA-α makes the clearest period-level claim: a plain vision-language model (VLM) plus a small MLP action head can match or beat heavier vision-language-action systems when the recipe is controlled. Its specialist model reports 98.8 average on LIBERO, 88.2 on RoboTwin 2.0 clean*, and 53.8 on RoboCasa-GR1, with the simple MLP head matching or beating more complex heads in the same setup. The ablations are as important as the headline numbers. Extra robot pretraining helps some benchmarks and hurts others, and common pipeline additions give only small gains once task data is large. This day’s strongest concrete result is not a new stack. It is a cleaner accounting of what actually changes outcomes.

Benchmarks are getting sharper about what a representation must do before it is called useful for control. LARYBench tests both semantic action decoding and low-level trajectory regression, and the result is blunt: general visual encoders are already stronger than many specialized latent action models. V-JEPA 2 reaches 76.62% average action classification, while DINOv3 posts the best control regression at 0.19 average MSE. AffordSim pushes on a different failure mode. It shows that imitation policies still break on tasks that depend on acting on the right object part. Across 17 representative tasks, Pi 0.5 averages 61%, but pouring and mug hanging remain far below easy grasping, and affordance-aware trajectory generation beats generic grasping by large margins on those tasks.

World models in this period try to make planning and action decoding more semantically grounded. AIM inserts a spatial value map between future prediction and action generation, then reports 94.0% average success on RoboTwin 2.0 Easy and 92.1% on Hard. Grounded World Model uses a shared language-image latent space for model-predictive control, so candidate futures are scored against text instructions instead of a goal image. On WISER, that gives 0.87 test success versus a 0.22 average for VLA baselines, while keeping train success at 0.92. DexWorldModel points to the same concern at deployment time: predict semantic features, keep memory fixed with horizon, and hide some inference behind robot execution. The latency claim is concrete, about 50%, but the task-performance evidence in the available excerpt is still thin.

Efficiency work is getting closer to robot-specific failure analysis. DA-PTQ treats quantization error as a control problem, with the key risk defined as trajectory drift over time rather than static layer reconstruction error. The method adds interface compensation and drift-aware mixed precision, aiming to keep low-bit VLA behavior close to full precision without extra runtime cost after calibration. The idea fits the day’s broader emphasis on execution fidelity, but the available excerpt does not include benchmark tables, so this remains a plausible systems contribution with limited visible verification in the local corpus.