Skip to content

agent_environments.md

Latest commit

 

History

History
93 lines (66 loc) · 3.35 KB

agent_environments.md

File metadata and controls

93 lines (66 loc) · 3.35 KB

Agent Environments

The AD agents evaluated in the HABIT paper have conflicting dependency requirements (different PyTorch versions, bundled libraries, etc.). This guide documents the separate conda environments needed to run each agent.

Why Separate Environments?

  • InterFuser and TransFuser require PyTorch 1.12.x with CUDA 11.3
  • BEVDriver requires PyTorch 2.0.x with a Llama backbone via the transformers library
  • These PyTorch versions cannot coexist in a single environment
  • The CARLA 0.9.14 PythonAPI is shared across all environments (added via PYTHONPATH)

Environment Overview

Environment Python PyTorch Use For
habit 3.8 none Core benchmark, NPC agent, custom agents
habit-interfuser 3.8 1.12.1+cu113 InterFuser (+ TransFuser, same deps)
habit-bevdriver 3.8 2.0.1 BEVDriver (LLM backbone)

Base Environment (habit)

The base environment runs the benchmark framework, NPC agent, and any custom agent that does not require a deep learning backbone.

conda env create -f environment.yml
conda activate habit

This is the only environment needed if you are:

  • Developing and testing a custom agent
  • Running the NPC or dummy baseline agents
  • Studying the evaluation framework and metrics

InterFuser / TransFuser (habit-interfuser)

Coming soon — Full setup instructions are being prepared.

Key details for early adopters:

  • Both InterFuser and TransFuser share the same dependency set (PyTorch 1.12.x, timm)
  • InterFuser bundles its own timm fork — do not install timm from pip
  • The original InterFuser targets CARLA 0.9.10.1; adaptation to 0.9.14 is required for HABIT
  • Pretrained weights need to be downloaded from the respective repositories
# Placeholder — full instructions coming soon
conda create -n habit-interfuser python=3.8
conda activate habit-interfuser
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
pip install -r requirements.txt
# + InterFuser-specific dependencies

BEVDriver (habit-bevdriver)

Coming soon — Full setup instructions are being prepared.

Key details for early adopters:

  • BEVDriver uses a Llama language model backbone via the transformers library (>=4.28)
  • Requires PyTorch 2.0.1 for compatibility with the LAVIS library (bundled in BEVDriver)
  • Pretrained weights need to be downloaded from the BEVDriver repository
# Placeholder — full instructions coming soon
conda create -n habit-bevdriver python=3.8
conda activate habit-bevdriver
pip install torch==2.0.1+cu117 torchvision==0.15.2+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
pip install transformers>=4.28
pip install -r requirements.txt
# + BEVDriver-specific dependencies

Running Agents

Regardless of which environment you use, the evaluation command is the same:

conda activate <environment-name>

export CARLA_ROOT=/path/to/carla
export PYTHONPATH=$CARLA_ROOT/PythonAPI/carla:$(pwd):$(pwd)/scenario_runner:$PYTHONPATH

bash scripts/run_evaluation.sh /path/to/agent.py

The --agent-config flag can be passed for agent-specific configuration:

bash scripts/run_evaluation.sh /path/to/agent.py --agent-config /path/to/config.yaml

See custom_agents.md for the full AutonomousAgent interface specification.