Humanoid Task Studio (HTS)

The open standard for training, benchmarking, and composing humanoid skills in simulation.

HTS is building the missing infrastructure layer in humanoid robotics: a standardized, open ecosystem for skill learning — from locomotion and manipulation to choreography and social gestures.

If large companies are building humanoid hardware,
HTS is building the open intelligence layer.

---

Why HTS?

There is no unified open standard for humanoid skill training.

HTS provides:

• Structured humanoid task packs (task.yaml)
• Reproducible RL + imitation learning baselines
• Standardized evaluation harness + metrics
• Composable high-level skill API
• Simulation-first, sim-to-real-ready architecture

Mission:
Make humanoid skill research accessible to anyone with a laptop.

---

Core Features

• Automatic task discovery via standardized task specification
• Dual control interface:
  • low_level joint control
  • high_level skill abstraction
• PPO and Behavior Cloning baselines
• Deterministic seeding for reproducibility
• CI-tested development workflow
• Extensible simulator abstraction (MuJoCo-first backend)

---

Roadmap

Planned milestones:

• Visual observation wrapper (RGB + depth)
• Sim-to-real hardware bridge (abstract interface + reference adapters)
• Vision-conditioned policy examples
• Expanded task packs (kitchen, dance, locomotion)
• Benchmark leaderboard export
• Multi-simulator backend support

---

Safety

HTS does not include combat or violent tasks.

Starter packs focus on non-violent locomotion, manipulation, and choreography.
See docs/safety_policy.md.

---

Install

Create a virtual environment and install in editable mode:

python -m venv .venv
source .venv/bin/activate   # Windows PowerShell: .venv\Scripts\Activate.ps1
pip install -e .[dev]
pre-commit install

---

Quickstart

Train a PPO policy on a starter task:

hts train task=starter_pack/pick_and_place algo=ppo

Evaluate a trained checkpoint:

hts eval task=starter_pack/pick_and_place checkpoint=artifacts/ppo_latest.pt

Run a scripted demo:

hts demo task=starter_pack/reach_object

---

Add a Task

Task packs live under:

src/hts_tasks/packs/<pack_name>/<task_name>/task.yaml

To create a new task:

1. Copy the task_template/ directory.
2. Modify task.yaml with task metadata and reward specification.
3. Optionally implement a scripted expert in expert_policy.py.
4. Add a matching config under configs/task/.

See docs/add_task_5min.md for a step-by-step guide.

---

Project Layout

src/
  hts_core/        # env API, simulator abstraction, metrics, wrappers, CLI
  hts_tasks/       # task packs, task loader, task logic
  hts_baselines/   # PPO and BC
  hts_datasets/    # demo format + converters
  hts_eval/        # evaluation harness and reports
  hts_bridge/      # sim-to-real bridge (in progress)
configs/           # Hydra configs
examples/          # end-to-end scripts
docs/              # documentation
tests/             # smoke and determinism tests

---

Reproducibility

• Deterministic seeds for numpy, torch, and Gym environments
• Hydra config snapshots per run
• JSON + Markdown evaluation reports

---

Contributing

Contributions are welcome.

See CONTRIBUTING.md for guidelines and task_template/ for adding new humanoid tasks.

---

🚀 Getting Involved

HTS is early and actively evolving.

If you want to contribute:

• Start with issues labeled good-first-issue
• Comment on an issue before opening a PR
• Follow CONTRIBUTING.md guidelines
• Join roadmap discussions in Issues

Areas especially welcome:

• New task packs
• Vision integration
• Sim-to-real adapters
• Benchmark experiments

---

License

Apache-2.0 (LICENSE)