World Models and Simulation

A three-layout Blender study separated camera occlusion from gripper approach clearance: 100 mm of clearance still left all 25 sampled sightlines blocked.

Export a Humanoid-Gym policy from Isaac Gym and replay it in MuJoCo to locate interface and physics mismatches. A second-engine pass is a screen, not hardware proof.

ASAP fits a residual from real humanoid rollouts, uses it to fine-tune a policy in simulation, then deploys the policy alone. Trace each evidence boundary before transfer.

DINO-WM predicts action-conditioned visual features and plans toward a goal image without reconstructing every candidate frame. Compare rollout accuracy and executed success together.

TACTO renders vision-based tactile observations while an external engine supplies contact dynamics. Choose tests that separate image handling from physical fidelity.

BEHAVIOR 2026 assigns public indices 0–9 to reported rollouts and 10–19 to pre-final testing. Freeze the policy, evaluator and timeout before reporting a score.

FPS matters only after the scene, cameras, image channels and omitted pipeline work match. Use a workload table to compare ManiSkill 3 visual throughput and memory.

When simulation and real trials rank robot policies differently, trace the first divergence through model inputs, action conversion and executed motion before selecting a checkpoint.

Spatial, Object, Goal and the 90-to-10 task boundary test different transfer claims. Preserve exposure, order and fixed starts before comparing policies.

The same robot command can give a world model different forecasts as an absolute target or a relative offset. Use paired inputs to test that failure before planning with it.