AgentSkillsCN

torch-sim-playground

当在PyTorch中搭建强化学习仿真平台时，可为环境、智能体、训练循环、日志记录与评估环节创建清晰的接口，并支持确定性种子的初始化。

SKILL.md
--- frontmatter
name: torch-sim-playground
description: Use when implementing an RL simulation playground in PyTorch. Create clean interfaces for env, agent, training loop, logging, and evaluation with deterministic seeding support.

Implement a minimal, testable RL playground in PyTorch.

Requirements:

  • •Define src/envs/ with a base Env interface: reset(seed) -> obs step(action) -> obs, reward, terminated, truncated, info
  • •Define src/agents/ with:
    • •model.py (torch.nn.Module)
    • •policy.py (action selection, exploration)
    • •learner.py (update step)
  • •Define src/train.py and src/eval.py as pure functions that take a config dict.
  • •Add scripts/train.py entrypoint:
    • •loads config
    • •sets seeds
    • •runs training
    • •writes logs to results/<run_id>/
  • •Add scripts/eval.py and scripts/make_figures.py placeholders.
  • •Add a tiny smoke test:
    • •run 100 steps on CPU, ensure no crash, logs created.

Logging:

  • •Save config.json
  • •Save metrics.jsonl or metrics.csv
  • •Save summary.json with final metrics

Determinism:

  • •Provide a single function set_seed(seed) used everywhere.