具身智能：端到端栈与数据闭环加速落地但操控仍是瓶颈
Key Questions
What progress is being made in embodied intelligence systems?
End-to-end demonstrations from Lanner AstraEdge, first-person datasets like Gen Ego Data, and visual memory layers from Memories.ai show the path to productization. Industrial cases include Skild AI and Visteon-NVIDIA.
How does GEM improve embodied AI capabilities?
GEM (Generative Supervision) embeds depth map generation into VLM pre-training, significantly boosting spatial and physical execution performance of embodied agents.
What are the remaining bottlenecks for embodied robots?
Fine-grained visual manipulation, long-term data loops, and production costs (target ~$20k per unit) continue to limit scale despite hardware and model advances.
What research unifies vision-language-action modeling for robots?
Qwen-VLA provides a unified framework across tasks, environments, and robot embodiments to improve generalization.
How is Wayve expanding beyond self-driving cars?
Wayve is launching an AI lab to apply its autonomous systems expertise to broader embodied intelligence applications.
端到端演示（Lanner AstraEdge）、第一视角数据集（Gen Ego Data）、视觉记忆层（Memories.ai）與工业化案例（Skild AI、Visteon‑NVIDIA）表明具身产品化路径正在形成。近期新增：GEM（Generative Supervision）通过将深度图生成嵌入 VLM 预训练，显著提升具身智能体的空间与物理执行能力。但视觉级精细操控、长期数据闭环与量产成本仍为关键瓶颈（行业目标约 $20k/台）。