Long-Term Memory for VLA-based Agents in Open-World Task Execution
A new AI framework uses long-term memory to master multi-step chemical experiments, outperforming existing models.
A research team led by Xu Huang has introduced ChemBot, a novel AI framework designed to solve a critical bottleneck in laboratory automation: executing long, complex chemical experiments. Current Vision-Language-Action (VLA) models, which enable robots to see, understand, and act, struggle with "long-horizon reasoning"—they can't remember past successes or failures across multi-step tasks, leading to repetitive trial-and-error. ChemBot addresses this with a dual-layer, closed-loop architecture. Its core is a "progress-aware" Skill-VLA model that performs hierarchical task decomposition, breaking down a high-level goal like "synthesize compound X" into actionable steps.
The system's breakthrough is a dual-layer memory architecture that consolidates successful action sequences into retrievable assets, allowing the robot to learn from experience. It uses a Model Context Protocol (MCP) server to efficiently orchestrate specialized sub-agents and tools. To further smooth execution, the team implemented a future-state-based asynchronous inference mechanism to prevent trajectory discontinuities. In extensive tests on collaborative robots, ChemBot demonstrated superior operational safety, precision, and task success rates compared to existing VLA baselines, proving its capability in open-world, long-horizon experimentation.
- Uses a dual-layer memory architecture to store and retrieve successful task trajectories, enabling learning from experience.
- Integrates a progress-aware Skill-VLA model with an MCP server for hierarchical task decomposition and efficient tool orchestration.
- Demonstrated on real robots, achieving higher safety, precision, and success rates in complex chemical protocols than current VLA models.
Why It Matters
This brings us closer to fully autonomous AI lab assistants that can reliably execute days-long experiments, accelerating scientific discovery.