Robot Learning Is Getting Cheaper to Judge, Not Just to Train

Humanoid control from egocentric data

In a preprint, HuGe is presented as a family of vision-language-action models designed for humanoid whole-body control, pre-trained on more than 2,000 hours of egocentric human and robot data. Its pipeline runs through egocentric pre-training, humanoid adaptation and downstream fine-tuning, constructs pseudo-proprioceptive states through simulation playback during pre-training, and modifies arm appearance in egocentric observations through re-rendering during fine-tuning. m:…6482

The authors report state-of-the-art performance on the HumanoidArena simulation benchmark, and say augmenting robot demonstrations with visually adapted egocentric demonstrations increased success on a pick-and-drop task for the Unitree G1 robot. As preprint results, those claims await independent replication, but the design bet is clear: human first-person experience, adapted to a robot body, is cheaper to collect than robot-only demonstrations at the same scale. m:…6482

Judging a model without running it

In a preprint, Sensorimotor Alignment, written as S squared, is proposed as a low-cost offline predictor of robot model performance, measuring how sensory representations match physical motion geometry. It compares neighborhoods in a model sensory representation space with neighborhoods of end-effector trajectories, on the idea that episodes near each other in sensory space should also sit near each other in motor space when alignment is high. m:…4420

The authors report that S squared correlates positively with downstream robot task success across model families, and that vision-language-action task success tracks the alignment of vision features rather than text features. The practical claim is that the metric can identify promising vision-language backbones before costly fine-tuning into action models, replacing some expensive closed-loop rollouts, which the authors note are the typical but costly way robot foundation models are evaluated. m:…4420

Planning from reusable causes

In a preprint, CausGround improves embodied planning by extracting reusable action preconditions and effects from interaction experience, organizing that causal knowledge in a factorized world-state representation capturing object attributes and relations between objects. Multimodal large language models often fail to capture environment-dependent preconditions and effects, the authors argue, while existing planning approaches demand costly environment-specific training or rely on unstructured interaction experience. m:…2733

The reported gains are large: average task success on VirtualHome rose from 24.9 percent to 53.3 percent, closed-loop planning success on RoboCasa rose from 43.0 percent to 57.0 percent, and interaction reasoning accuracy on SWITCH rose from 25.8 percent to 31.9 percent, all without retraining the underlying model. The system reasons backward from interpreted goals to derive executable action sequences and forward from current observations, and it can expand its reusable causal knowledge without retraining, which is the mechanism behind the claimed transfer across environments. m:…2733

The deployment side still needs a referee

May Mobility offers the counterweight to preprint optimism: Edge Case completed the first phase of an independent safety assessment of the company safety case framework, applying its DevSafeOps methodology to evaluate structure and completeness and alignment with UL 4600 and the Open Safety Case standard. Edge Case found the framework consistent with industry best practices, and May Mobility is streamlining the framework for easier interpretation and management while formalizing release decisions to include independent assessment. m:…7210

The company is also strengthening safety performance measurement against a human-driving benchmark and expanding monitoring and evaluation of real-world driving data, after completing more than half a million commercial autonomous rides. The lesson that ties the four items together is that robot progress now depends as much on cheap judgment, reusable knowledge and independent review as on bigger training runs. m:…7210