Zhejiang University proposes EMem-Bench for embodied agent memory
Zhejiang University's OmniAI team introduced EmbodiedMemory-Bench (EMem-Bench), a benchmark with four task types—passive observation, dynamic tracking, interaction failure and experience generalization—and 2,554 executable samples testing whether embodied agents can use past observations, state changes and action feedback in later tasks. The team also proposed an external memory system, Embodied-Memorizer (EMem), and fine-tuned Qwen3-VL-8B into EMem-8B; in the authors' evaluation, EMem raised average success from 35.3% to 58.9% with a GPT-5.4-mini backbone. Under the paper's full-context setting, Gemini-3-Flash led at 64.2% average success and GPT-5.4 reached 56.8%. The report says the benchmark drew nearly 30,000 Hugging Face downloads in about three weeks, with results mainly from simulation.
The authors' analysis of 100 failure trajectories found 91% involved memory-related problems, indicating memory—not just perception and planning—is a key bottleneck for long-horizon embodied tasks.
新智元 2026-10-18 13:00 北京
新智元报道
设想一个日常场景：
家庭机器人刚刚在茶几上看见一本书，去厨房完成几项任务后，接到取书的指令，却又开始挨个房间搜索。它已经试过一个打不开的抽屉，过了一会儿，还是回到同一个抽屉前反复尝试。
对人来说，「记得东西放在哪里」「知道它后来被挪走了」「记住哪些办法行不通」，都是日常生活中的基本能力。
#Zhejiang University#OmniAI#EmbodiedMemory-Bench#Embodied-Memorizer#EMem-8B#Qwen3-VL-8B#Gemini-3-Flash#GPT-5.4-mini#Hugging Face#AI2-THOR#ProcTHOR