具身推理与任务规划 / Embodied Reasoning and Task Planning¶
English contract¶
- Objective: convert an instruction and world state into a typed, inspectable plan that can execute skills, observe outcomes, and replan after failure.
- Inputs: task instruction, grounded entities, available skills, preconditions/effects, safety constraints, and feedback events.
- Stages: parse → ground → plan → validate → execute skill → observe → replan or terminate.
- Acceptance: measure parse validity, grounding accuracy, subgoal completion, replan count, constraint violations, and final task success on held-out scenarios.
- Evidence: the included rule-based path is interface-tested; it does not validate general language-model reasoning. Apply the validation policy.
目标与边界¶
把自然语言目标转成可执行、可监控、可重规划的技能序列。规划器只负责决定“做什么”;技能控制器、世界状态估计和安全监督决定“是否能安全做到”。
前置知识与输入¶
Pipeline¶
| 阶段 | 关键动作 | 输出/检查 |
|---|---|---|
| 1. Grounding | 解析实体、关系和约束 | grounded goal |
| 2. Task planning | 分解 locate/approach/grasp/move/place | typed TaskPlan |
| 3. Validation | 检查技能存在、参数合法、前置条件 | executable plan |
| 4. Skill execution | 调用感知与控制技能 | skill status/telemetry |
| 5. Monitoring | 判断完成、失败、超时和安全事件 | structured feedback |
| 6. Replanning | 更新世界状态并恢复或重规划 | revised plan |
| 7. Audit | 保存输入、计划、工具调用与结果 | trace 与错误分类 |
运行与产物¶
python scripts/run_pipeline.py --run embodied-reasoning
入口:rule_based_planner.py。当前示例验证语言到 TaskPlan/SubGoal 的结构化接口;它不执行机器人技能。
验收门槛¶
- 计划只能调用白名单技能,参数必须通过 schema 验证。
- 每个 subgoal 有前置条件、成功条件、超时与失败恢复策略。
- 在含糊指令、对象不存在、技能失败和场景变化时测试。
- 报告计划解析成功率、子目标完成率、重规划次数和最终任务成功率。
常见失败:LLM 生成不存在的技能、只生成文本不生成类型化计划、执行失败后盲目重试、计划日志无法复现。