Pouring policy for Alloy (Pour-v0)
A policy for Pour-v0, a ready-made environment of the Alloy physics engine: a
stream of 100 ml of water falls and the policy moves a bowl under it. The water is simulated with Alloy's liquid
solver.
| File | What it is |
|---|---|
pour_v0.pt |
the checkpoint at 38.5 min of training (iteration 10) |
evaluate.py |
runs the policy on Pour-v0 and counts successes |
Result
118 of 128 episodes (92.2 %) put at least 75 % of the water in the bowl, for the checkpoint at 38.5 min of training. Training success first passed 75 % at 24.9 min.
Measured with alloy-engine 0.1.0rc2 from PyPI (torch 2.15, warp-lang 1.18.0, mujoco 3.17.0) on one NVIDIA L4 (about 23 GB; a GCP spot g2-standard-8, driver 580.178.04), from one training run.
- Task: Alloy's shipped
Pour-v0. An episode succeeds when at least 75 % of the water is in the bowl at the end. - Training: Alloy's rsl_rl wrapper with its default settings (PPO, a 128 x 128 MLP), 512 worlds.
- Evaluation: 128 episodes, seed 1000.
Run it
python3 -m venv .venv && source .venv/bin/activate # Python 3.11 or 3.12
pip install "alloy-engine[envs]" rsl-rl-lib
python -c "from huggingface_hub import snapshot_download; snapshot_download('argonrobotics-ai/alloy-pour', local_dir='alloy-pour')"
cd alloy-pour
python evaluate.py
The first run compiles kernels once; the evaluation of 128 worlds then takes about 6 minutes on an L4. Liquid is slow on a CPU: use a GPU for this one.
More on pouring in Alloy: how-to.
Notes
- The number above is from alloy-engine 0.1.0rc2. Later versions can give a different count, and on a GPU two runs of the same evaluation can differ by a few episodes (worlds in contact do not repeat bit for bit on a GPU).
pour_v0.ptis an rsl_rl checkpoint (actor and critic weights and the observation normaliser), saved bytorch.save;evaluate.pyloads it with rsl-rl-lib (tested with 5.5.2).- Licence: Apache-2.0, for the weights and the code on this page.
Alloy is a GPU physics engine for robotics: code, documentation, training guide, paper.