SPEAR: Realistic Embodied AI Simulator
adobe-research · hf · 2026-07-17
SPEAR is a simulator for realistic embodied AI research, functioning primarily as a Python library that connects to and programmatically controls any Unreal Engine application via a modular plugin architecture.
The paper highlights several key improvements:
- Exposes 14K+ UE functions via Python, making it easier to program than existing UE simulators.
- Directly renders realistic 1920×1080 images at 73 FPS.
- Outputs ground-truth modalities typically unavailable in existing UE simulators, such as non-diffuse intrinsic image decomposition, material IDs, and physically-based shading parameters.
The authors also propose a high-level programming model that organizes UE workloads into graphs with arbitrary data dependencies, executing deterministically within a single UE frame. The paper showcases various use cases: controlling multiple embodied agents with different action spaces, generating city-scale realistic scenes, manipulating procedural content generation systems, synchronizing multi-view face renders, conducting interactive co-simulation with MuJoCo, and editing scenes via natural language using AI coding assistants.
More from Embodied
- Tesla’s summer update lets Grok make calls, control climate, and play music — Polymarket · 2026-07-22
- Hugging Face Teases Agentic Training Environments with OpenEnv for August Launch — mervenoyann · 2026-07-22
- NVIDIA shows 22 SIGGRAPH papers and Omniverse tools for robot simulation — facontidavide · 2026-07-22
- Nothing phone mockup turns a film joke into a modular design meme — ZeYanjie · 2026-07-22
- Lightwheel AI Launches SimReadyGen: Text-to-Physics-Accurate Robot Sim Assets — ZeYanjie · 2026-07-22
- Humanoid robot sorting packages in a warehouse sparks debate over job loss — MonaJalal_ · 2026-07-22