Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery
FOSA discovers which tasks a robot can learn. A GFlowNet over symbolic task skeletons proposes a task, an LLM compiles it into a MuJoCo Playground environment, and PPO trains a task policy on it: the policy's learning curve is the GFlowNet's reward. The task policies are then distilled into one language-conditioned student.
conda create -n fosa python=3.11 -y
conda activate fosa
conda install conda-forge::uv -y
uv pip install -e . --no-cacheThe robot models are downloaded from the dataset on first use.
LLM
FOSA talks to any OpenAI-compatible chat endpoint. The paper uses Gemma 4 31B, which you can serve locally, e.g. with vLLM:
vllm serve google/gemma-4-31B-it --served-model-name gemma-4-31B-it --port 8000or use a hosted model: awesome-free-llm-apis lists free ones.
export FOSA_LLM_ENDPOINT=http://localhost:8000/v1/chat/completions # default
export FOSA_LLM_MODEL=gemma-4-31B-it # default
export FOSA_LLM_API_KEY=... # if the server needs oneTask grammars of the paper
Each robot's task grammar is written once by the LLM. To reuse the paper's, copy them into the log directory:
mkdir -p logs/gemma-4-31B-it && cp -r data/grammars/* logs/gemma-4-31B-it/- Bump mujoco version to 4
- Full VLA ckpts for all robots
Important
The paper runs each robot and method on 4 GPUs (NVIDIA A100, 40 GB).
Robots: go2 (quadruped), jackal (vehicle), booster (humanoid), inspire_hand (hand).
FOSA
python -m fosa.discovery.fosa --robot go2 --gpu-ids 0,1,2,3Ablations
python -m fosa.discovery.fosa --robot go2 --gpu-ids 0,1,2,3 --method fosa_rnd # FOSA-RND
python -m fosa.discovery.fosa --robot go2 --gpu-ids 0,1,2,3 --method fosa_llm # FOSA-LLMBaselines
python -m fosa.discovery.mcts --robot go2 --gpu-ids 0,1,2,3 --method mcts # also llm_mcts, muzero (MCTS-Learn)
python -m fosa.discovery.curriculum --robot go2 --gpu-ids 0,1,2,3 # LLM-CLEach run trains 256 tasks; run the same command again to resume an interrupted run.
To run every robot and method of the paper automatically, one job each:
scripts/slurm.sh go2,jackal,booster,inspire_hand all 256python -m fosa.learning.dagger --robot go2 --method fosa --gpu-ids 0,1,2,3On Slurm (the paper uses seeds 1, 2 and 3):
FOSA_STUDENT_SEED=1 scripts/slurm_student.sh go2,jackal,booster,inspire_hand allFOSA's students from the paper are in checkpoints/<robot>/student.pt.
python -m fosa.evaluation.heldout prepare go2 # compile the held-out tasks
python -m fosa.evaluation.heldout train go2 # train a reference policy per task
python -m fosa.evaluation.heldout evaluate go2 --checkpoint checkpoints/go2/student.pt@inproceedings{yu2026generative,
title={Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery},
author={Youwei Yu and Lantao Liu},
booktitle={10th Annual Conference on Robot Learning},
year={2026},
url={https://openreview.net/forum?id=weyfPS89NJ}
}