feat(training): release V0.9.1 避障训练与基础策略迁移
web-platform-ci / TypeScript, lint, unit, build (push) Has been cancelled
web-platform-ci / Playwright E2E (push) Has been cancelled

This commit is contained in:
2026-09-08 10:50:13 +08:00
parent fa5485049a
commit 438e56bcc8
113 changed files with 15027 additions and 539 deletions
+5 -2
View File
@@ -12,7 +12,8 @@ from .schema import validate_proposal
SYSTEM_PROMPT = """你是 Unitree Go2 强化学习奖励调参专家。
只根据提供的数值配置、训练曲线摘要和固定评估结果提出下一轮稀疏修改。
必须优先保持速度跟踪与跌倒安全门槛;每轮最多修改四个白名单标量,不得改变符号、函数、传感器或结构。
必须优先保持当前任务的客观安全门槛:Flat速度跟踪/跌倒,Obstacle无碰撞到达/跌倒。
每轮最多修改四个白名单标量,不得改变符号、函数、传感器、评估协议或结构。
不要建议 Python 代码、命令、文件路径或白名单外参数。输出必须符合 RewardProposal schema。
"""
@@ -101,7 +102,9 @@ class DeepSeekAdvisor:
result = self._agent().run_sync(prompt)
output = result.output
patch = validate_proposal(
{"weights": dict(output.weights), "params": dict(output.params)}, previous
{"weights": dict(output.weights), "params": dict(output.params)},
previous,
task_id=context.get("task", "Unitree-Go2-Flat"),
)
try:
usage = result.usage()