EVENT DOSSIER
The researchers proposed Continual Field-Adaptive Models (CFAMs) aimed at solving the problem where machines in critical task domains cannot retain their existing capabilities when faced with novelty after deployment. This model adopts a complementary learning architecture, consisting of a frozen slow component (including three cortical layers for perception, reasoning, and action) and a fast-learning Capsule Field, supporting small-sample skill installation in the lab and autonomous updates without gradients after deployment. CFAM was evaluated in five forms: robotic arms, quadruped robots, humanoid robots, drones, and off-road vehicles. The results showed that it could achieve equivalent operating levels with only 40% of standard full training data (or 2.5 times the number of trajectories); during testing, the autonomous capture of verified near-distribution-out cases increased the success rate of actions by 13.9 percentage points, and negative transfer in sequence simulations was -0.5 percentage points, significantly better than the -11.4 percentage points of the LoRA method.
研究人员提出 Continual Field-Adaptive Models (CFAMs)，旨在解决关键任务领域机器在部署后面对新颖性时无法保留既有能力的难题。该模型采用互补学习架构，包含一个冻结的慢速组件（含感知、推理、行动三个皮层）和一个快速学习的 Capsule Field，支持实验室内的少样本技能安装及部署后的无梯度自主更新。CFAM 在机械臂、四足机器人、人形机器人、无人机和越野车辆五种形态上进行了评估，结果显示其仅需标准全量训练数据的 40%（或 2.5 倍轨迹数）即可达到同等操作水平；在测试时，对验证过的近分布外案例的自主捕获使动作成功率提升 13.9 个百分点，且序列仿真中的负向迁移为 -0.5 个百分点，显著优于 LoRA 方法的 -11.4 个百分点。