📚 @saturdayrobotic Robotics & World Model Reading Club 14, @deeptechvc Deep Tech Week Recap SPEAR: Programmable Photorealistic Simulation at Scale — 14K Functions, 150 MP/s Rendering & Infrastructure for Controllable World Models @mikeroberts3000 (@AdobeResearch) presented SPEAR: How do we programmatically control Unreal Engine for photorealistic, fully controllable synthetic data? 🎮 Synthetic data is essential for vision/MM foundation models, enabling instance segmentation (Hypersim, SAM/SAM2), intrinsic decomposition (TOG'24), 3D/4D reconstruction (VGGT), surface normals (CVPR'24 Oral), monocular depth (CVPR'24), and 3D detection (CVPR'24). ⚠️ Existing UE interfaces fall short: Blueprints are expressive but visual-only; the official ~14K-function Python API cannot control active simulations, cannot run outside the editor, has limited rendering. ⚡ SPEAR introduces a production Python interface with 14K Blueprint-equivalent APIs, 150 MP/s rendering, active-simulation control, outside-editor execution, frame-scoped begin_frame()/end_frame() synchronization, async futures for non-blocking execution, native-frame-rate Python, and >15× UnrealCV+ performance. 💡 Open discussion & Possible directions • JoyAI-VL-Interaction → Cosmos3: pilot-assisted bootstrapping toward controllable, real-time/full-duplex world models with active intervention instead of passive prediction. • ImageWAM: image foundation models as the backbone of action-conditioned World Action Models, learning directly in image space while SPEAR provides controllable scene state, embodiment, physics and feedback. • Cosmos Policies: deploy policies through Blueprint-equivalent APIs instead of narrow observation-action interfaces, enabling richer closed-loop reasoning and execution. • LoopWAM: equilibrium/fixed-point world-model inference with iterative refinement (>1000 refinement loops continue improving), where each refinement immediately informs simulator actions. • Orthogonal refinement: refinement loops are independent of temporal rollout—reasoning improves at every timestep without waiting for long-horizon simulation, dramatically increasing efficiency. 🎤 Panel, moderated by @bcristei (@SHACK15sf) Simone T. (Saturn Dynamics): compute-efficient world models; observability, verifiability, scalability; evaluation gap; precision gap; whether scaling data alone can close it. @shumochu (@gi_labs): end-to-end Physical AI infrastructure—stereo egocentric headsets, UMI grippers, motion-capture gloves, global data ops, post-training and physical RL—turning human motion into sovereign robot policies. Margaret Zhang (ThirdbrainLabs): infrastructure for the next 1B specialized models, enabling robotics/industrial teams to continuously own, customize and improve private Physical AI models by encoding proprietary operational knowledge with far lower data, labor and training cost.