Research Track 02 · Physical Interaction

通用移动物理 AI 的物理交互安全 Physical Interaction Safety for General-Purpose Mobile Physical AI

通用移动物理 AI 一旦进入真实公共空间，安全问题就不只是“撞没撞上”。更关键的是：它如何接近人、避让人、接触人、搬运物体、处理地面差异，以及理解人的意图和边界。

Once general-purpose mobile physical AI enters real public spaces, safety is not simply about whether it collides. The harder questions are how it approaches people, avoids people, makes contact, carries objects, handles floor variability, and interprets human intent and boundaries.

1. 传统机器人安全给了什么基础 1. What Traditional Robot Safety Already Provides

ISO 10218 和 ISO/TS 15066 给工业机器人和协作机器人提供了非常重要的基础，包括速度限制、力限制、保护停止、协作空间、接触力阈值等。ISO 13482 则进一步面向个人护理机器人，把人体接触、移动辅助和服务场景纳入考虑。

ISO 10218 and ISO/TS 15066 provide important foundations for industrial and collaborative robots, including speed limits, force limits, protective stops, collaborative workspaces, and contact-force thresholds. ISO 13482 extends the discussion toward personal care robots, including human contact, mobility assistance, and service scenarios.

但通用移动物理 AI 的问题更开放。它可能在家庭、商场、医院、校园、园区、仓库、酒店和街区之间切换，面对的是非专业用户、不规则地面、不可预测人群和自然语言任务。

General-purpose mobile physical AI is more open-ended. It may move across homes, shopping malls, hospitals, campuses, industrial parks, warehouses, hotels, and street blocks, while interacting with non-expert users, irregular floors, unpredictable crowds, and natural-language tasks.

2. 物理交互安全至少有五类风险 2. At Least Five Risk Classes Matter

风险类型	典型场景	可测指标	与 SOTIF 的关系
接近风险	机器人从侧后方接近行人、儿童或老人	最小距离、相对速度、TTC、让行时机	感知和行为预测不足可能导致无故障危险
接触风险	机械臂递送物品、清洁、开门、辅助搬运	接触力、接触位置、持续时间、人体部位敏感度	功能目标正确但动作策略不安全
稳定性风险	斜坡、门槛、湿滑地面、电梯缝隙、负载偏移	倾覆裕度、制动距离、轮地附着、负载摆动	环境条件超出系统隐含能力边界
任务对象风险	搬运热饮、药品、刀具、玻璃、易碎品	对象分类置信度、抓取稳定性、掉落概率	任务理解和对象风险建模不足
社会交互风险	插队、阻挡通道、误入儿童活动区、过度靠近陌生人	空间侵入、路径礼让、用户拒绝信号识别	语义和社会规范理解不足会转化为安全风险

Risk Class	Typical Scenario	Measurable Indicators	Connection to SOTIF
Approach Risk	A robot approaches a pedestrian, child, or older adult from the side or rear	Minimum distance, relative speed, time-to-collision, yielding timing	Insufficient perception or behavior prediction can create hazardous behavior without a component fault
Contact Risk	A manipulator hands over an object, cleans a surface, opens a door, or assists with carrying	Contact force, contact location, duration, sensitivity of the body region	The functional goal may be correct while the action strategy is unsafe
Stability Risk	Slopes, thresholds, wet floors, elevator gaps, or shifted payloads	Tip-over margin, braking distance, wheel-ground adhesion, payload oscillation	The environment may exceed the system’s implicit capability boundary
Task-Object Risk	Transporting hot drinks, medicine, knives, glassware, or fragile items	Object-classification confidence, grasp stability, probability of dropping	Task understanding and object-risk modeling may be insufficient
Social Interaction Risk	Cutting through queues, blocking passages, entering children’s activity areas, or standing too close to strangers	Space intrusion, courteous path behavior, recognition of user rejection signals	Insufficient understanding of semantic and social norms can become a safety risk

3. 为什么移动物理 AI 需要跨本体指标 3. Why Mobile Physical AI Needs Cross-Embodiment Metrics

传统机器人标准往往默认一个相对清晰的本体：机械臂、AGV、服务机器人、个人护理机器人。但通用移动物理 AI 很可能是移动底盘、双臂、夹爪、升降机构、语音交互和视觉语言模型的组合。

Traditional robot standards often assume a relatively clear embodiment: a manipulator, an AGV, a service robot, or a personal care robot. General-purpose mobile physical AI may instead combine a mobile base, dual arms, grippers, lifting mechanisms, voice interaction, and vision-language models.

这意味着评价指标不能只绑定某一种机械结构。更合理的做法是从任务和风险出发：接近、接触、搬运、避让、让行、解释用户指令、识别拒绝信号、处理异常环境。

Metrics therefore should not be bound to one mechanical structure. A more robust approach starts from tasks and risks: approach, contact, carrying, avoidance, yielding, instruction interpretation, rejection-signal recognition, and handling of abnormal environments.

4. 运行时安全层不能只靠急停 4. Runtime Safety Cannot Rely Only on Emergency Stop

急停是必要的，但它是最后一道防线。真正的机器人 SOTIF 架构需要在运行时持续判断：当前任务是否仍在边界内，当前人机距离是否可接受，当前动作是否需要降速、绕行、暂停、请求确认或交给人。

Emergency stop is necessary, but it is the last line of defense. A Robot SOTIF architecture must continuously judge whether the current task remains within boundary, whether the human-robot distance is acceptable, and whether the current action should slow down, reroute, pause, request confirmation, or hand control to a human.

这类运行时安全层应该至少包括环境异常检测、人体接近监控、动作约束、负载状态监控、语义任务校验和回退策略。

Such a runtime safety layer should include, at minimum, environmental anomaly detection, human-proximity monitoring, action constraints, payload-state monitoring, semantic task validation, and fallback strategies.

5. 可执行的研究路线 5. A Practical Research Agenda