Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/facebookresearch/HyperAgents/llms.txt

Use this file to discover all available pages before exploring further.

The Genesis domain evaluates agents that generate reward functions for reinforcement learning controllers. The agent is given a task description and must write a Python reward function that, when used to train an RL policy, produces the desired locomotion behavior in the Genesis physics simulator.

​
What It Evaluates

Genesis tests reward function engineering ability. The agent’s output (a reward function) is not judged directly; instead, it is used to train an RL policy via rsl-rl, and the resulting policy’s behavior is evaluated in simulation. The primary metric is average_fitness — a normalized 0–1 score measuring how well the trained policy executes the task.

​
Evaluation Setup

  • Total simulation time: 20 seconds per evaluation
  • Episode duration: 4.0 seconds (200 steps at dt = 0.02 s)
  • Parallel environments: 4096 simulated simultaneously
  • Fitness score range: 0 (worst) to 1 (best)
  • RL training: 101 policy update iterations before evaluation
  • Early termination: an episode ends early if the robot falls (roll or pitch > 10°)

​
The Three Environments

Go2WalkingCommand-v0 — The Unitree Go2 robot must learn to walk forward at a commanded speed.
  • Task: Go2WalkingCommand-v0/speed
  • Linear velocity range: [0.2, 0.8] m/s in the x direction
  • Default episodes: 6
  • Domain: genesis_go2walking

​
Requirements

Genesis requires a CUDA-compatible GPU. PyTorch must be installed before Genesis, and the correct PyTorch version depends on your CUDA version.

​
Install PyTorch

Check your CUDA version first: 
nvidia-smi
Then install the matching PyTorch version following the official guide.

​
Install Genesis

pip install --upgrade pip
pip install git+https://github.com/Genesis-Embodied-AI/Genesis.git
pip install rsl-rl-lib==2.2.4 tensorboard==2.20.0

​
Setup and Run

1

Run the initial evaluation

python -m domains.harness \
  --domain genesis_go2walking \
  --run_id initial_genesis_go2walking_0 \
  --num_samples 3
2

Generate the report

python -m domains.report --domain genesis_go2walking \
  --dname ./outputs/initial_genesis_go2walking_0

​
num_workers Constraint

Genesis must be run with --num_workers 1. The harness enforces this automatically — regardless of the value passed via --num_workers, the Genesis branch in harness.py hardcodes num_workers = 1 before composing the Hydra config:
# domains/harness.py
elif "genesis" in domain:
    num_workers = 1  # forced
    cfg = compose(
        config_name="config",
        overrides=[
            f"eval.num_workers={num_workers}",
            ...
        ]
    )
This is because Genesis runs GPU-accelerated physics simulation that cannot safely be parallelized across threads.

​
Hydra Configuration

The config file is at domains/genesis/config/config.yaml. Key sections: 
envs:
  names: "go2walking"
  num_envs: 4096

rl_trainer:
  max_iterations: 101

rl_eval:
  max_steps: 1000  # 20 seconds at 0.02 s/step
  record_video: False

eval:
  num_workers: 1
  num_episodes:
    go2walking: 6
    go2walkback: 6
    go2hop: 6

​
Output Structure

Outputs are written to <output_dir>/<env_name>/<task_name>/ and include:
  • chat_history_*.md — agent conversation log
  • rl_eval_<episode_idx>/ — evaluation results (JSON log, eval_100.mp4 video)
  • rl_train_<episode_idx>/ — training artifacts (model checkpoints model_0.pt, model_100.pt, TensorBoard events, config pickle)

​
Domain Properties

PropertyValue
Score keyaverage_fitness
Splitstrain only
Eval subsetfull dataset
Ensemble supportedNo
Staged eval samples3 out of 6 (50%)
num_workersAlways 1

Build docs developers (and LLMs) love

Get started for freeTalk to us