Robot policies are judged by object identity, contact, and simulator fidelity

VLA policies are adding explicit object and relation structure inside the action path. OA-WAM separates a fixed object address from changing object content, which makes target binding testable through slot interventions. It reports 97.8% average success on LIBERO and a swap-binding cosine of 0.87, while holistic baselines stay at 0.09 or lower.

TriRelVLA takes a related route through object, hand, and task nodes. Its graph uses four relation types: task-object, task-hand, object-hand, and object-object. The available excerpt gives mechanism detail and claimed gains for cross-scene, cross-object, and cross-task manipulation, but no quantitative benchmark table.

World-model work is being evaluated by whether its latent state preserves action-relevant information. One study trains action-conditioned diffusion world models on BridgeV2 while varying only the latent encoder. Semantic encoders such as V-JEPA 2.1 and SigLIP 2 beat reconstruction-focused VAEs on policy rollouts, action recovery, and success classification. V-JEPA 2.1_96 reaches 0.362 consensus success in DiT-S rollouts, compared with 0.169 for VAE.

EA-WM attacks the action-conditioning side. It converts robot kinematics into camera-aligned visual fields, including arm skeletons, gripper geometry, end-effector heatmaps, and pose axes. On WorldArena, it reports 76.60 P3CScore versus 71.08 for CogVideoX, with ablations showing losses when kinematic-to-visual fields or event-aware fusion are removed.

Human data appears in two forms: large video pretraining and small motion sets refined into executable dexterity. HumanNet reports 1,000,000 hours of human-centric video with first-person and third-person views, captions, motion descriptions, and hand/body signals. In a controlled VLA validation, 1,000 hours of egocentric HumanNet pretraining reportedly matches or slightly beats 100 hours of real-robot CoBot pretraining under the same LingBot-VLA setup.

DexSynRefine shows the smaller-data path. It starts with seven human-object interaction demonstrations per task, expands them to about 300 trajectories per task, and adds residual reinforcement learning plus contact and dynamics adaptation. Across five dexterous tasks, task-space residual actions reach 68.1% mean simulated success, while raw kinematic retargeting stays near failure.

Two papers focus on whether robot policy claims hold under distribution changes. VLA-GSE uses parameter-efficient fine-tuning: it trains 114.04M of 4,551.85M parameters, or 2.51%, with shared and routed low-rank experts. It reports 81.2% average zero-shot success on LIBERO-Plus, ahead of full fine-tuning and several PEFT baselines in its same-backbone comparison.

VISER tests the evaluation side. It builds 1,049 physically based rendering assets across 319 categories and adds specular materials, soft shadows, and reconstructed real-world tasks. The benchmark reports an average Pearson correlation of 0.92 between simulation and real-world policy performance, with visual cues changing task outcomes sharply: eggplant-in-pot success rises from 10% without specular highlights to 90% with them.