microduck_rl
microduck_rl 性能与成本：测量 rollout、物理与复核
基准吞吐、显存、并行环境、检查点和策略审核。
你将学会
- Map environment, simulator, and runner
- Run a reproducible simulation
- Identify sim-to-real gates
开始前需要
- Basic Git and command-line usage
- Comfort reading a project README
You can explain the project, run its documented first step, and decide what to verify before adopting it.
先看结论
- Throughput depends on physics and environment count.
- Accepted-policy cost includes evaluation and review.
- Pruning must preserve lineage and safety evidence.
Freeze the workload
固定任务、仿真器、环境数、物理步和评估协议，记录 steps/s、p50/p95、显存、重置、崩溃、存储与审核时间。
Use the pinned source evidence, record versions and assumptions, and ask a qualified reviewer to validate the result before adopting it.
Measure complete outcomes
Record steps/second, p50/p95 episode time, GPU/CPU memory, utilization, reset rate, crashes, checkpoint size, video cost, and reviewer minutes. Compare policies on equal disturbances and seeds.
Optimize carefully
Parallel environments improve throughput until memory or simulator contention dominates. Early stopping and checkpoint pruning save storage, but never remove the evidence needed to reproduce a safety decision.
Use the pinned source evidence, record versions and assumptions, and ask a qualified reviewer to validate the result before adopting it.
如何选择
| 比较维度 | 方案 A | 方案 B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
实施步骤
- 1
Freeze task, simulator, seeds, and device.
- 2
Measure rollout, physics, reset, and memory metrics.
- 3
Compare equal evaluation protocols.
- 4
Set budget and regression thresholds.
可复制示例
envs x steps -> throughput/p50/p95
GPU/CPU + resets + crashes -> run cost
checkpoint + disturbance eval + review -> accepted policy常见问题
Is steps/second enough?
No. Include episode success, stability, resets, memory, evaluation, and review time.
Can more parallel environments hurt?
Yes. Contention, memory pressure, and unstable physics can offset throughput gains.
资料来源
- microduck_rl repository来源核查 2026-08-31