Downloads · 30 days
3
0% of all-time downloads
CursedRock17/gazebo-gymnasium-policies
gazebo-gymnasium-policies is a reinforcement learning model from CursedRock17. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for stable-baselines3.
Trained policies for gazebogymnasium (v0.1.0, commit 51ebdd0) — Gazebo Harmonic + ROS 2 Jazzy robots wrapped as Gymnasium VecEnvs, trained headlessly with Stable-Baselines3 via the in-process backend (no launch, unthr…
Downloads · 30 days
3
0% of all-time downloads
All-time downloads
605
Public
Repo size
46.3 MB
Likes
0
Public
Click a slice to open those files.
.zip21.5 MB · 100%
From the Hugging Face model README
Trained policies for gazebo_gymnasium
(v0.1.0, commit 51ebdd0) — Gazebo Harmonic + ROS 2 Jazzy robots wrapped as
Gymnasium VecEnvs, trained headlessly with Stable-Baselines3 via the
in-process backend (no launch, unthrottled real_time_factor).
Most runs below are a single, unswept probe — establishing learnability, not necessarily hitting the repo's stricter per-env "solved" bar. CartPole, line_follower, inverted_double_pendulum, hopper, and walker2d are the exceptions: all were swept (research-informed hyperparameters where noted) and fully solved. See each subfolder's model card for exact hyperparameters and the achieved metric against that bar.
To reuse a policy in the Gazebo GUI (pixi run sim + pixi run deploy --backend harness):
hf download CursedRock17/gazebo-gymnasium-policies <env>/model.zip --local-dir models/
| Env | Algo | n_agents | Timesteps | Result | vs. solved bar |
|---|---|---|---|---|---|
| cartpole | PPO | 16 | 250k (swept, 4 configs) | 500.0 / 500.0 (reward/len) | SOLVED (500-step cap) |
| cartpole_continuous | SAC | 16 | 200k | 500.0 / 500.0 | SOLVED (500-step cap) |
| inverted_double_pendulum | PPO | 16 | 300k (swept) | 4357.3 reward, 10/10 eval episodes at cap | SOLVED (1000-step cap) |
| hopper | PPO | 16 | 600k (swept + resumed) | 1002.2 / 999 (reward/len), 10/10 eval episodes | SOLVED (1000-step cap, reward ≥1000) |
| walker2d | PPO | 16 | 600k (swept + resumed) | 1522.9 / 999 (reward/len), 10/10 eval episodes | SOLVED (1000-step cap, reward ≥1500) |
| half_cheetah | PPO | 16 | 600k (swept + resumed) | ~3.01 m/s sustained (median 3.25) | open-ended — not yet at MuJoCo-solved (~5-6 m/s), real progress from a 1.46 m/s probe |
| reacher | PPO | 16 | 100k | −12.4 / 50.0 | ≥−5 bar not yet met; a later SAC probe reached −8.48, still the best known |
| line_follower | PPO+CNN | 4 | 400k (GPU, swept + DR) | 300.0 / 300.0, 40/40 episodes | SOLVED (300-step cap), DR-tuned for sim-to-real |
Evaluated with 10 deterministic episodes per model on the single-env
gymnasium.make(env_id) path (seeded, headless in-process), separately from
the training run itself — except line_follower, evaluated across 8 agents
(40 episodes) since DR makes a single-agent eval too high-variance to trust;
see its model card. inverted_double_pendulum/hopper/walker2d were evaluated
across 8 agents (80 episode-instances) for the same reason a single eval
episode's ep-length report can be too noisy to trust.
Ant is spec'd and mechanically verified (100% survival under random
actions) but is not yet a solved or even progressing policy: both PPO and
SAC converge to a stand-still local optimum (0.0 m/s sustained, alive the
whole episode) rather than walking — a reward-shaping issue (alive_bonus),
not an algorithm-choice one. No checkpoint uploaded since neither actually
walks. See docs/examples/README.md in the main repo.