train

AI Community

Repo stars 1,155
License BSD-2-Clause
Author repo LearningHumanoidWalking

Interpretation is structured for decision-making; original keeps the upstream SKILL.md unchanged.

Decide Fit First

Core job: Launch a training run for a robot environment using PPO Parse the user's request from $ARGUMENTS and construct a training comman…
Best fit: Use it when the task has reusable inputs, steps, and validation criteria rather than a one-off answer.
Avoid forcing it: If the source lacks commands, platform support, or external-service evidence, keep those fields unknown instead of guessing.

Design Intent

Structure: The skill is organized around “Command Template”, “Available Environments”, “Hyperparameters (defaults)”, “Instructions”, showing how the author expects the agent to judge fit, collect context, and produce verifiable output.
Trigger evidence: Prioritize the author’s wording around when to use it, what context to collect, and what output shape to produce.
Evidence boundary: Author text states facts, repository files prove commands and paths, and Fluxly only adds fit, limits, and usage judgment.

How To Use It

Inputs: Provide target material, scope, expected result, forbidden changes, and validation method.
Invocation: Name train directly; if the source includes slash commands, start with the command and then add task context.
Validation: Start small and check whether the result follows “Command Template / Available Environments / Hyperparameters (defaults)” before expanding.

Boundaries And Review

Dependencies: It usually needs no extra API key, so start with a small validation task.
Permissions: Declared permissions include read / write; ask the agent to state file, command, and rollback boundaries before acting.
Quality bar: A useful result names the deliverable, evidence, and next action. Generic prose means the task needs tighter context.

/train — Launch a PPO Training Run

Parse the user's request from $ARGUMENTS and construct a training command.

Command Template

RAY_ADDRESS= uv run python run_experiment.py train --env <ENV> --logdir <LOGDIR> [OPTIONS...]

Available Environments

Name	Description
`cartpole`	Cartpole swing-up (simplest, good for testing)
`h1`	Unitree H1 standing task
`jvrc_walk`	JVRC humanoid basic walking
`jvrc_step`	JVRC humanoid stepping with planned footsteps

Hyperparameters (defaults)

Flag	Default	Description
`--n-itr`	20000	Training iterations
`--lr`	1e-4	Learning rate
`--gamma`	0.99	Discount factor
`--std-dev`	0.223	Action noise
`--learn-std`	off	Learn action noise (flag)
`--entropy-coeff`	0.0	Entropy regularization
`--clip`	0.2	PPO clipping
`--minibatch-size`	64	Minibatch size
`--epochs`	3	Optimization epochs per update
`--num-procs`	12	Parallel workers
`--num-envs-per-worker`	1	Vectorized envs per worker
`--max-grad-norm`	0.05	Gradient clipping
`--max-traj-len`	400	Episode horizon
`--eval-freq`	100	Eval every N iterations
`--seed`	None	Random seed
`--device`	auto	Training device (auto/cpu/cuda)
`--no-mirror`	off	Disable symmetry wrapper (flag)
`--recurrent`	off	Use LSTM policy (flag)
`--continued`	None	Path to pretrained weights

Instructions

Determine the environment name from the user's request. If ambiguous, ask.
Use --logdir /tmp/training_runs unless the user specifies a different path.
Only include flags that differ from defaults — keep the command clean.
Show the user the full command you're about to run.
Run the command in the background using run_in_background: true on the Bash tool. Set a generous timeout (600000ms).
After launching, tell the user the logdir path and how to check progress (you can tail the output using the task ID).
If the user asks to check on training, use TaskOutput with block: false to check the latest output.

Cartpole-Specific Defaults

For cartpole, these settings are known to work well with the current defaults (--lr 3e-4 --max-grad-norm 0.5 --lam 0.95 --gamma 0.99):

--minibatch-size 256
--std-dev 0.15 --learn-std --entropy-coeff 0.01
--max-traj-len 500 --n-itr 500 --num-procs 12
--no-mirror (cartpole has no body symmetry)

Suggest these defaults when the user trains cartpole, but let them override.

Fluxly profile Author and license come from source; runtime, permissions, and network are Fluxly detections or estimates

Fluxly category: AI
Author-declared agents: No explicit declaration found; this is not inferred or tested compatibility
Static check: 94 / 100 · heuristic scan, not runtime safety proof
Author / version / license: @rohanpsingh · BSD-2-Clause
Fluxly token estimate: Lean
Fluxly setup estimate: Plug-and-play
External API key: No requirement detected
Detected OS requirements: Unspecified
Runtime requirements: Python
Detected file/system behavior: Read-only
Write / modify
Detected network behavior: Local-only
Install commands: None (reference only)

Profile is derived at build time from SKILL.md and install vectors. Subject to drift from author intent.

Heads up: 未限定 allowed-tools，默认拥有全部工具权限。

Output preview train.preview

The current SKILL.md does not define a fixed output example.

SKILL.md SKILL.md

# /train — Launch a PPO Training Run

Parse the user's request from `$ARGUMENTS` and construct a training command.

## Command Template

```
RAY_ADDRESS= uv run python run_experiment.py train --env <ENV> --logdir <LOGDIR> [OPTIONS...]
```

## Available Environments

| Name | Description |
|------|-------------|
| `cartpole` | Cartpole swing-up (simplest, good for testing) |
| `h1` | Unitree H1 standing task |
| `jvrc_walk` | JVRC humanoid basic walking |
| `jvrc_step` | JVRC humanoid stepping with planned footsteps |

## Hyperparameters (defaults)

| Flag | Default | Description |
|------|---------|-------------|
| `--n-itr` | 20000 | Training iterations |
| `--lr` | 1e-4 | Learning rate |
| `--gamma` | 0.99 | Discount factor |
| `--std-dev` | 0.223 | Action noise |
| `--learn-std` | off | Learn action noise (flag) |
| `--entropy-coeff` | 0.0 | Entropy regularization |
| `--clip` | 0.2 | PPO clipping |
| `--minibatch-size` | 64 | Minibatch size |
| `--epochs` | 3 | Optimization epochs per update |
| `--num-procs` | 12 | Parallel workers |
| `--num-envs-per-worker` | 1 | Vectorized envs per worker |
| `--max-grad-norm` | 0.05 | Gradient clipping |
| `--max-traj-len` | 400 | Episode horizon |
| `--eval-freq` | 100 | Eval every N iterations |
| `--seed` | None | Random seed |
| `--device` | auto | Training device (auto/cpu/cuda) |
| `--no-mirror` | off | Disable symmetry wrapper (flag) |
| `--recurrent` | off | Use LSTM policy (flag) |
| `--continued` | None | Path to pretrained weights |

## Instructions

1. Determine the environment name from the user's request. If ambiguous, ask.
2. Use `--logdir /tmp/training_runs` unless the user specifies a different path.
3. Only include flags that differ from defaults — keep the command clean.

…

Evidence boundary and execution chain

Author text anchors workflow facts; Fluxly only indexes current sections, terms, files, and commands.

sections -> Command Template → Available Environments → Hyperparameters (defaults) → Instructions → Cartpole-Specific Defaults
terms    -> Parse the user's request from $ARGUMENTS and construct a training command. · 1. Determine the environment name from the user's request. · Suggest these defaults when the user trains cartpole, but let them override.
files/cmd -> $ARGUMENTS · cartpole · jvrcwalk · jvrcstep · --n-itr · --lr · --gamma · --std-dev
body sha256 -> 67a84e6e2fb7