Unilab Train

UniLab Train

Monitor and manage UniLab RL training jobs on local WSL Ubuntu 24.04 (CPU only).

Start Here

Read references/platforms.md before issuing any command that depends on:

• WSL invocation template
• Log paths (training log + run dir)
• Checkpoint paths (WSL + Windows mirror)
• Resume semantics (max_iterations is additional, not total)
• Viewer commands (viser / MuJoCo native / TensorBoard)
• Common gotchas (HFENDPOINT mandatory, saveinterval=50, viser rgba bug, etc.)

Treat references/platforms.md as the source of truth for platform-specific paths and commands.

Operating Rules

• All commands run via wsl -d Ubuntu -u u20174 -- bash -c "..." wrapper.
• The wrapper prints a Chinese garbage warning on every invocation — ignore it, never try to "fix" it.
• Always prefix HF_ENDPOINT=https://hf-mirror.com for any command that might trigger HF download.
• Prefer compact metric snapshots over raw log dumps. Offer raw dump only if user asks.
• Health assessment: provide a one-line Chinese verdict + concrete next action.
• Never kill training without explicit user confirmation.
• For resume: use algo.load_run=, NEVER resume=true (UniLab has no such flag).

Slash-Flag Invocation

User may invoke via /unilab-train with optional flags. Interpret the flags as intent:

| Flag | Action | |---|---| | (none) or --status | Quick health snapshot — is training alive, current iter, ETA, 1-line verdict | | --tail | Detailed tail of latest training log + metric trend analysis | | --metrics | Full metric table extraction (all reward components + losses + curriculum) | | --best | Find best ckpt via tfevents parse, identify nearest saved model*.pt | | --resume | Resume training from a saved ckpt (remind about maxiterations semantics) | | --viser [port] | Start viser on given port (default 8080); report URL | | --mjc | Start MuJoCo native viewer via WSLg (real-time GLFW window) | | --tb | Start TensorBoard on port 6006 if not already running | | --record | Record Motrix demo mp4 (playenvnum=1, calculate play_steps from motion length) |

If user passes an unknown flag, treat it as conversational intent and ask for clarification.

Core Tasks

Check connectivity

wsl -d Ubuntu -u u20174 -- bash -c "echo OK && uptime"

Check training status (--status)

Detect from process list:

wsl -d Ubuntu -u u20174 -- bash -c "ps aux | grep -E 'train|play_viser|play_interactive|tensorboard' | grep -v grep"

Report:

• training running or not
• PID and elapsed time
• algo (ppo / sac / appo)
• task and sim
• active log file path
• background services (viser / TB) alive?

If nothing is running but log file mtime is recent (5 min stale while process supposedly running, suspect silent crash or WSL restart — verify with ps.

Tail log (--tail)

Primary use case. Read active training log and extract structured metrics.

Step 1 — find latest training log:

wsl -d Ubuntu -u u20174 -- bash -c "ls -t /home/u20174/UniLab/logs/*.log 2>/dev/null | head -3"

If user named the run explicitly (e.g. g1_walk_migrated_resume.log), use that. Otherwise pick the most recent.

Step 2 — extract last ~50 lines containing key metrics:

wsl -d Ubuntu -u u20174 -- bash -c "tail -200 /home/u20174/UniLab/logs/ | grep -E 'Learning iteration|Mean reward|Mean action std|Mean episode|Steps per|Collection time|Learning time|ETA|reward/tracking|reward/alive|reward/feet_phase|reward/penalty|reward/pose|curriculum/|Mean value|Mean surrogate|Mean entropy'"

Step 3 — present compact snapshot (NOT raw dump):

== Training:  ==
iter X/Y | elapsed HH:MM:SS | ETA HH:MM:SS
throughput: N steps/s | collection: X.Xs | learning: X.Xs

== Reward ==
Mean reward: CURRENT (trend: was A iter N iters ago → now B, Δ)
  tracking_lin_vel:    X.XX    ← main forward signal
  tracking_ang_vel:    X.XX
  feet_phase:          X.XX
  alive:              XX.XX    ← survival bonus
  penalty_action_rate:-XX.XX   ← should decrease as std drops
  penalty_orientation:-X.XX
  penalty_ang_vel_xy: -X.XX
  pose:               -X.XX

== Policy ==
action_std: X.XX (was 1.0 at start, target 0.2-0.4)
value_loss: X.XX | surrogate: -X.XX | entropy: X.XX | lr: X.XXe-X

== Curriculum ==
penalty_scale: X.XX (range 0.5-1.0)
avg_episode_length: X.X (max = max_episode_steps)

Step 4 — one-line Chinese health verdict + next action:

• 健康: "Reward 持续上涨，action_std 收敛中，按 ETA X 小时完成。无需干预。"
• 警告: "Episode length 卡在 X，curriculum 未升级。可能需要降噪声或放宽 termination。"
• 危险: "Reward 不涨反跌 / actionstd 卡在 1.0 / learningtime >5s。建议停训调参。"
• 中断: "训练 log X 分钟没更新，进程已死。WSL 可能重启。需要 resume from model_XXX.pt。"

Full metrics extraction (--metrics)

Same as --tail but show full metric table including all sub-rewards, losses, and curriculum state. Useful for comparing runs.

Find best checkpoint (--best)

wsl -d Ubuntu -u u20174 -- bash -c "cd /home/u20174/UniLab && /home/u20174/.local/uv run --no-sync python ', size_guidance={'scalars': 0})
ea.Reload()
events = ea.Scalars('Train/mean_reward')
best = max(events, key=lambda e: e.value)
print(f'Best iter: {best.step}, reward: {best.value:.4f}')
# also print top 5 for context
top5 = sorted(events, key=lambda e: -e.value)[:5]
for e in top5:
    print(f'  iter {e.step}: {e.value:.4f}')
EOF"

Then find nearest saved ckpt (multiples of 50):

wsl -d Ubuntu -u u20174 -- bash -c "ls /home/u20174/UniLab/logs/rsl_rl_ppo///model_*.pt | sort"

Pick model_.pt. Copy to Windows:

wsl -d Ubuntu -u u20174 -- bash -c "cp /home/u20174/UniLab/logs/.../model_XXXX.pt /mnt/d/Desktop_Files/UniLab/checkpoints///model_XXXX.pt"

Resume training (--resume )

wsl ... train --algo ppo --task  --sim  \
  algo.load_run= \
  algo.max_iterations= training.no_play=true \
  > /home/u20174/UniLab/logs/_resume.log 2>&1

Critical: warn user that max_iterations is additional, not total. If they want total 3000 starting from iter 500, set max_iterations=2500.

What gets restored: network weights, optimizer state, normalization stats, action_std, curriculum state, iter counter. What resets: env RNG (acceptable for PPO).

viser (--viser)

wsl ... uv run python scripts/play_viser.py \
  task=/mujoco \
  interactive.action_mode=policy \
  algo.load_run= \
  viser.port= \
  > /home/u20174/UniLab/logs/viser_.log 2>&1 &

Open http://localhost: in Windows browser. Ctrl+Shift+R to refresh.

If colors look wrong: viser reads geom_rgba directly, not material rgba. Edit XML to add explicit rgba="..." on each geom.

MuJoCo native viewer (--mjc)

wsl ... bash -c "export DISPLAY=:0 && uv run python scripts/play_interactive.py \
  --algo ppo --task  --sim mujoco \
  algo.load_run= \
  interactive.action_mode=policy"

Opens GLFW window via WSLg. Materials work correctly (no rgba hack). Controls: Space=pause, N=step, +/-=speed, Esc=quit.

If window immediately closes with "Done." in log, DISPLAY may be unset — explicitly export DISPLAY=:0.

TensorBoard (--tb)

wsl ... uv run tensorboard --logdir /home/u20174/UniLab/logs --port 6006 --bind_all \
  > /home/u20174/UniLab/logs/tb.log 2>&1 &

Open http://localhost:6006. Restart TB after starting a new training run — its data server doesn't auto-discover new subdirs.

Record Motrix demo (--record)

wsl ... eval --algo ppo --task  --sim motrix \
  --render-mode record \
  training.play_env_num=1 training.play_steps= \
  algo.load_run=

Output: src/unilab/assets/checkpoints//play_video.mp4. Copy to Windows:

wsl ... bash -c "cp .../play_video.mp4 /mnt/d/Desktop_Files/UniLab/videos//motrix/model_0.mp4"

Known motion durations (frames @ 50Hz):

• dance: 870 (17.4s)
• wallflip: 198 (4.0s)
• boxtracking: 725 (14.5s)

Response Style

Order:

1. Current state (alive? iter X/Y? ETA?)
2. Key metrics (reward + main components)
3. Health verdict (one-line Chinese)
4. Recommended next action

Keep concise unless user asks for full raw logs or full metric tables.

Common Issues Quick Reference

| Symptom | Likely cause | Fix | |---|---|---| | --load-run must be '-1' or dir name | CLI flag rejects paths | Use Hydra override algo.load_run= | | Could not override 'resume' | UniLab has no resume flag | Just algo.load_run= (no resume=true) | | HF download RuntimeError: client closed | hfhub 1.18.0 bug + GFW | HF_ENDPOINT=https://hf-mirror.com prefix | | uv sync --extra viser removes motrix | uv sync matches exactly | uv sync --extra viser --extra motrix | | TB not showing new run | data server caches subdir list | Restart TB process | | viser color wrong | rgba bug | Add explicit rgba="..." to geom XML | | MuJoCo viewer exits immediately | DISPLAY not propagated | export DISPLAY=:0 before uv run | | Motrix mp4 only 1.6s | playsteps too small | play_steps = seconds × 50 | | Reward stuck negative | Strict termination + no alive bonus | Tune task yaml: maxtilt, alive, curriculum | | episodelength stuck low | Termination too strict | Raise maxtiltdeg, lower minbaseheight | | Learning time >5s | CPU thermal throttle | Reduce num_envs or wait for cooldown |

Project Variants

G1 walk_flat (PPO)

Focus on:

• Mean reward trend (positive = good)
• trackinglinvel (main signal, target >0.8)
• episode_length (target >500 for stable walking)
• penaltyactionrate (should drop below -10 by iter 500)
• curriculum penaltyscale (should ramp 0.5 → 1.0 as episodelength grows)
• action_std (target 0.2-0.4 after 500 iter)

Go2 joystick_flat (PPO)

Similar but:

• Lower maxepisodeseconds (typically 10s = 500 steps)
• Faster convergence (smaller observation space)
• curriculum not always enabled

G1 motion tracking (Motrix)

Focus on:

• motion reconstruction error
• per-joint tracking accuracy
• play_steps must match motion file length

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: phanghonghao
• Source: phanghonghao/THU-Awesome-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.