qualia

# Qualia

Fine-tune and iterate on robotic foundation models — VLAs, reward models, and more — on cloud GPUs.

## Setup

```bash
export QUALIA_API_KEY="your-api-key"
```

## Quick Commands

```bash
# Account
{baseDir}/scripts/qualia.sh credits               # Check credit balance
{baseDir}/scripts/qualia.sh instances             # GPU options and pricing

# Models & data
{baseDir}/scripts/qualia.sh models                             # VLA types and camera slot requirements
{baseDir}/scripts/qualia.sh dataset-keys <huggingface/dataset> # Image keys for camera mapping

# Projects
{baseDir}/scripts/qualia.sh projects                       # List your projects
{baseDir}/scripts/qualia.sh project-create "My Project"   # Create a project
{baseDir}/scripts/qualia.sh project-delete <project_id>   # Delete a project

# Training
{baseDir}/scripts/qualia.sh hyperparams <vla_type> [model_id]  # Default hyperparams (model_id required for smolvla/pi0/pi05)
{baseDir}/scripts/qualia.sh finetune <project_id> <vla_type> <dataset_id> <hours> '<camera_json>'
{baseDir}/scripts/qualia.sh status <job_id>                # Training progress and phase history
{baseDir}/scripts/qualia.sh cancel <job_id>                # Stop a running job
```

## Launching a Fine-Tune Job

### 1. Pick a model

```bash
{baseDir}/scripts/qualia.sh models
```

Supported VLA types: `act`, `smolvla`, `pi0`, `pi05`, `gr00t_n1_5`, `sarm`

### 2. Check your dataset's image keys

```bash
{baseDir}/scripts/qualia.sh dataset-keys your-org/your-dataset
```

### 3. Map image keys to camera slots

Each VLA type has required camera slot names (shown in `models`). Build a JSON mapping:

```json
{"image_0": "front", "image_1": "wrist"}
```

### 4. Create a project (if needed)

```bash
{baseDir}/scripts/qualia.sh project-create "My Robot"
```

### 5. Launch

```bash
# smolvla/pi0/pi05 require --model
{baseDir}/scripts/qualia.sh finetune \
  <project_id> \
  pi0 \
  your-org/your-dataset \
  4 \
  '{"cam_1": "observation.images...

# Qualia Agent Skill 🤖🦾

Make **robotics training** a skill for your AI agent. This skill lets any agent fine-tune robotic foundation models on [Qualia](https://qualiastudios.dev) — launch training jobs, monitor progress, iterate on parameters, and build full pipelines from conversation.

## What can your agent do with this?

- Fine-tune robotic foundation models (ACT, SmolVLA, π0, π0.5, GR00T N1.5, and more)
- Monitor training progress in real-time
- Iterate on hyperparameters and reward functions
- Build full training pipelines from conversation
- Use Reward-Aware Behavior Cloning (RA-BC) with SARM reward models

## Supported Models

| Type | Description |
|------|-------------|
| `act` | Action Chunking Transformer — fast, lightweight |
| `smolvla` | SmolVLA — efficient open-source model |
| `pi0` | π0 — Physical Intelligence foundation model |
| `pi05` | π0.5 — dexterous manipulation variant |
| `gr00t_n1_5` | GR00T N1.5 — NVIDIA humanoid foundation model |
| `sarm` | SARM — reward model for RA-BC |

More models coming soon.

## Setup

1. Get a Qualia API key from the [Qualia app](https://app.qualiastudios.dev/)
2. Set the environment variable:
   ```bash
   export QUALIA_API_KEY="your-api-key"
   ```

### OpenClaw

Copy this skill to your workspace:
```bash
cp -r qualia-agent-skill ~/.openclaw/workspace/skills/qualia
```

Or install via ClawHub (coming soon):
```bash
clawhub install qualia
```

Then add `"qualia"` to your agent's skills list in `openclaw.json`.

### Claude Code / Other Agents

The `SKILL.md` file is a self-contained instruction set. Drop it into any agent's context that supports tool use and shell access.

## How it works

The skill provides a bash script (`scripts/qualia.sh`) that wraps the Qualia API. Your agent reads `SKILL.md` to understand the available commands and uses them to manage training jobs.

Example flow:
1. Agent checks available models → `qualia.sh models`
2. Inspects dataset image keys → `qualia.sh dataset-keys org/dataset`...