feat: release v1.0.1 CADWorld 网站与 LeKiwi 智能抓放
web-platform-ci / Standalone decision service (no cloud credentials) (push) Has been cancelled
web-platform-ci / TypeScript, lint, unit, build (push) Has been cancelled
web-platform-ci / Playwright E2E (push) Has been cancelled
lekiwi-compatibility / cpu-compatibility (push) Has been cancelled
web-platform-ci / Standalone decision service (no cloud credentials) (pull_request) Has been cancelled
web-platform-ci / TypeScript, lint, unit, build (pull_request) Has been cancelled
web-platform-ci / Playwright E2E (pull_request) Has been cancelled
lekiwi-compatibility / cpu-compatibility (pull_request) Has been cancelled
web-platform-ci / Standalone decision service (no cloud credentials) (push) Has been cancelled
web-platform-ci / TypeScript, lint, unit, build (push) Has been cancelled
web-platform-ci / Playwright E2E (push) Has been cancelled
lekiwi-compatibility / cpu-compatibility (push) Has been cancelled
web-platform-ci / Standalone decision service (no cloud credentials) (pull_request) Has been cancelled
web-platform-ci / TypeScript, lint, unit, build (pull_request) Has been cancelled
web-platform-ci / Playwright E2E (pull_request) Has been cancelled
lekiwi-compatibility / cpu-compatibility (pull_request) Has been cancelled
集成同源 BYOK 会话隔离、精简模型设置、官方订阅入口和 HTTPS 发布运维;保留本地训练/调参与控制能力。同步 npm 版本及 CHANGELOG,记录公网真实 API 验收仍待用户凭据。
This commit is contained in:
@@ -1,4 +1,17 @@
|
||||
# 自定义训练与前视射线避障
|
||||
# 统一强化学习训练控制台
|
||||
|
||||
## LeKiwi / LeKiwi Bundle
|
||||
|
||||
1. 主工作台导入机器人 ZIP,在 URDF 选项选对应的「移动操作训练场景」。也可先加载旧 LeKiwi v1 profile,再从控制台选择任务;开始训练会自动加载抓取/搬运场景。
|
||||
2. 启动配置了 `--mobile-python` 的训练服务([独立依赖配置](../training_server/README.md#lekiwi-一键训练与-go2-共用控制台)),在「控制台 → 强化学习任务」连接。面板按当前机器人选择任务,也允许手动切换。
|
||||
3. 默认先选「底盘接近」,配置示教初始化步数(0 可禁用)、位置随机化、独立评估回合,再配置迭代数、环境数(1–64)、每环境采样步数、设备、种子及物体/目标坐标。控制周期固定0.02s,Go2 地形/射线/基础策略/自调参配置隐藏。点击「发起本地训练」自动同步场景并创建后台作业,**不再下载训练包**。
|
||||
4. 共享日志、Loss/平均奖励曲线、进度与停止按钮。完成后点击「导入策略」自动拉取 ONNX 与 deployment,核验形状、任务、变体及 SHA-256,在当前主视口开始推断。暂停/重置/移除策略沿用主控制台。
|
||||
|
||||
5. 根据独立评估决定续训:点击「接续此作业(保留权重)」可同阶段继续。至少10回合、≥80%成功率、无安全终止才允许依次升级到末端接近、抓取放置;场景资产与已评估初态分布必须匹配。未达标策略可以调试导入,但不是成品策略。控制采用 v2 限速增量目标,旧68维 ONNX 会拒绝加载。详见 [训练课程与实测结果](../docs/mobile-training-curriculum.md)。
|
||||
|
||||
两种变体不能互用策略。移动操作任务使用专用物体/目标/地面,不使用 Go2 地图编辑配置;总采样步数为 `迭代 × 环境数 × 每环境采样步数`。物理在 CPU、PPO 网络可选 CPU/单 GPU;短训练只验证运行链路,不保证抓取收敛。独立 `/mobile.html` 产品页已移除;保留主工作台 RL、`tuning.html` 与 `/physics/mobile.html` 夹具。新增的 [LeKiwi 智能抓放](../docs/lekiwi-agent.md) 使用独立任务契约,切回训练会重建 RL 场景,不把双支撑台任务作为原训练快照。普通作业状态在服务重启后不保留,与现有 Go2 一致。
|
||||
|
||||
## Go2 自定义训练与前视射线避障
|
||||
|
||||
1. 启动仓库本地训练服务,在「控制台 → 强化学习任务」输入令牌并连接。
|
||||
2. 选择「前视射线避障导航」,设置地形、种子、障碍物参数、FOV、探测/安全距离及避障奖励权重,发起训练。界面显示进度、最近价值/策略/熵损失和原始日志。
|
||||
|
||||
Reference in New Issue
Block a user