Train Robotic AI Models using Qualia

Qualia

Fine-tune Vision-Language-Action (VLA) models for robotics on cloud GPUs.

Install and setup

Install from ClawHub / OpenClaw

Use the short slug:

openclaw skills install qualia-skill

If @fabbe1999/qualia-skill or fabbe1999/qualia-skill fails, retry with qualia-skill. Some OpenClaw installer versions only resolve the short slug even though the public ClawHub URL is https://clawhub.ai/fabbe1999/qualia-skill.

Configure API key

1. Sign up at app.qualiastudios.dev
2. Create an API key (Settings -> API Keys)
3. Set the env var for local shell use:

   export QUALIA_API_KEY="your-api-key"
   

For OpenClaw, prefer configuring the key in ~/.openclaw/openclaw.json so agent runs receive it:

{
  skills: {
    entries: {
      qualia: {
        enabled: true,
        env: {
          QUALIA_API_KEY: "your-api-key"
        }
      }
    }
  }
}

Notes for agents:

• The skill name is qualia, even though the ClawHub install slug is qualia-skill.
• openclaw skills install places files in the workspace skills/ folder. Start a fresh agent turn/session after installing so skills are reloaded.
• If the skill is installed but inactive, check skills.entries.qualia.enabled is not false and QUALIA_API_KEY is available.

Verify your install

First thing after setup, run the self-test:

python3 {baseDir}/scripts/qualia.py doctor

It checks the API key, auth/connectivity (/v1/credits), and the models endpoint. Exit 0 means everything works. Add --json for machine-readable results.

Machine-readable output

Every command accepts a global --json flag. In JSON mode, stdout carries exactly one JSON object or array and nothing else. Errors are emitted on stdout as {"error": {"code": , "message": "...", "details": ...}} with the matching exit code.

python3 {baseDir}/scripts/qualia.py --json credits
# {"balance": 90784}

Prefer --json when driving the CLI programmatically; parse stdout, branch on exit code.

Exit codes

When Someone Asks to Train a Model

They probably won't give you everything upfront. Here's what you need and how to get it:

1. Dataset: ask for their HuggingFace dataset ID (e.g. your-org/your-dataset)
2. Model type: if they don't specify, run models and help them choose:
   • Quick prototyping → suggest ACT (fast, no base model needed)
   • Production quality → suggest π0.5 or π0
   • Humanoid robots → suggest GR00T N1.5
   • Resource-conscious → suggest SmolVLA
3. Training duration: if unspecified, suggest 2 to 4 hours for a first run
4. Camera mapping: run dataset-keys on their dataset, then models to see required slots, and map them automatically. Confirm with the user before launching.

If the user already has a project, use it. Otherwise create one.

When Things Go Wrong

Always run status and share the full phase history with the user when debugging.

Quick Start

# See what models are available (always check, new ones are added regularly)
python3 {baseDir}/scripts/qualia.py models

# Check GPU options and pricing
python3 {baseDir}/scripts/qualia.py instances

# Check your credit balance
python3 {baseDir}/scripts/qualia.py credits

Train a Model

# 1. Discover image keys in your dataset
python3 {baseDir}/scripts/qualia.py dataset-keys your-org/your-dataset

# 2. Create a project
python3 {baseDir}/scripts/qualia.py project-create "My Robot"

# 3. Launch training
python3 {baseDir}/scripts/qualia.py finetune   your-org/your-dataset 4 \
  '{"cam_1": "observation.images.top"}' \
  --model  \
  --name "My run"

# 4. Monitor
python3 {baseDir}/scripts/qualia.py status 

Notes:

• Run models first to see which VLA types require --model and which don't
• Camera mappings map model slots (from models) to dataset image keys (from dataset-keys)
• Smart camera mapping: The API returns generic slot names (cam_1, cam_2, cam_3) but the underlying models have a specific input order. Map semantically using these known orders:
  • π0.5 / π0: cam_1 = base/overview camera, cam_2 = left wrist/arm, cam_3 = right wrist/arm
  • GR00T N1.5: cam_1 = base/overview camera, cam_2 = left wrist/arm, cam_3 = right wrist/arm
  • ACT / SmolVLA: cam_1 = primary camera, cam_2/cam_3 = secondary views
  • Fuzzy-match dataset keys to these roles: context_camera or base_0 → cam_1; left_wrist ≈ left_arm → cam_2; right_wrist ≈ right_arm → cam_3
• Omit --model for types that don't support custom models
• Estimate cost before launching: run instances to get credits/hr, multiply by hours. Tell the user the estimated cost before confirming.
• Dataset IDs on HuggingFace are case-sensitive, double-check the exact ID

Manage Jobs & Projects

python3 {baseDir}/scripts/qualia.py projects                     # List projects and jobs
python3 {baseDir}/scripts/qualia.py status               # Job status and phase history
python3 {baseDir}/scripts/qualia.py cancel               # Cancel a running job
python3 {baseDir}/scripts/qualia.py project-delete   # Delete a project

Custom Hyperparameters

# Get defaults
python3 {baseDir}/scripts/qualia.py hyperparams  [model_id]

# Validate overrides
python3 {baseDir}/scripts/qualia.py hyperparams-validate  '{"learning_rate": 1e-4}'

# Use in training
python3 {baseDir}/scripts/qualia.py finetune ... --hyper-spec '{"learning_rate": 1e-4, "num_epochs": 50}'

Finetune Flags

RA-BC (Reward-Aware Behavior Cloning)

Use a trained SARM reward model to weight training samples. Supported on smolvla, pi0, pi05.

python3 {baseDir}/scripts/qualia.py finetune \
   pi0 your-org/your-dataset 4 \
  '{"cam_1": "observation.images.top"}' \
  --model lerobot/pi0 \
  --rabc your-org/sarm-reward-model \
  --rabc-image-key observation.images.top \
  --rabc-head-mode sparse

Job Phases

queuing → credit_validation → instance_booting → instance_activation → instance_setup → dataset_preprocessing → training_running → model_uploading → completed

Terminal: completed, failed, cancelled

Live Docs

For the latest models, endpoints, and capabilities, always check the live documentation:

• LLM context: docs.qualiastudios.dev/llms.txt
• API reference: dev-docs.qualiastudios.dev/api/reference
• SDK: docs.qualiastudios.dev/sdk/overview
• Guides: docs.qualiastudios.dev/global/guides

Links

• Platform: https://app.qualiastudios.dev
• Docs: https://docs.qualiastudios.dev
• API: https://api.qualiastudios.dev (auth via X-API-Key header)