Closing the Reality Gap: Trustworthy Simulation and Synthetic Data for Embodied AI
具身智能中的高可信仿真与合成数据生成
| Chair: |
Co-chair: |
 |
 |
| Yan Xia |
Haowen Wang |
| University of Science and Technology of China, China |
Anhui University, China |
| |
|
| Summary: |
|
- General-purpose robots are moving rapidly toward real-world deployment, yet data remains a key bottleneck for embodied intelligence. Real-world data collection is costly, long-tail scenarios remain insufficiently covered, and safety validation is difficult to conduct at scale. As a result, building trustworthy, controllable, and scalable simulation environments that faithfully approximate real-world physical dynamics and interactions while enabling reliable policy generalization has become a central challenge in the field. This session focuses on trustworthy simulation and synthetic data generation, covering frontier topics such as differentiable physics, neural rendering, foundation-model-driven scene generation, and world models. We invite researchers from academia and industry to share recent advances and engineering practices, and to discuss how simulation and synthetic data can support model training, capability evaluation, and sim-to-real transfer for embodied AI, ultimately accelerating the deployment of general-purpose robots in real-world scenarios.
|
| |
|
- 通用机器人正在加速走向真实世界，但数据仍是制约具身智能发展的关键瓶颈。真实数据采集成本高、长尾场景覆盖不足，安全验证也难以大规模开展。如何构建可信、可控、可扩展的仿真环境，使其既能逼近真实物理与交互过程，又能支撑机器人策略在真实世界中的可靠泛化，已成为具身智能领域的重要问题。本专题聚焦高可信仿真与合成数据生成，关注可微分物理、神经渲染、大模型驱动的场景生成与世界模型等前沿方向，邀请学术界与工业界研究者分享最新进展与工程实践，共同探讨仿真与合成数据如何支撑具身智能模型训练、能力评估与真实场景迁移，加速通用机器人在真实场景中的落地应用。
|