V2.0.0
Bumblebee v2.0.0
Bumblebee v2.0.0 is a major release focused on non-deterministic, human-like mouse and keyboard automation powered by reinforcement learning.
This release ships a packaged SAC mouse policy, a cleaner runtime API, better keyboard controls, and a documented RL training workflow.
Highlights
- Packaged default RL mouse model:
sac_mouse_v2.zip - Runtime mouse paths can now come from the bundled RL policy or your own provider.
- New mouse movement profiles:
default,precise,fast,natural,messy. - New keyboard profiles:
default,fast,careful,messy,developer. - Expanded mouse helpers for clicking, dragging, scrolling, bounds-based targeting, and random points inside UI regions.
- Expanded keyboard helpers for hotkeys, navigation, editing, typing, paste fallback, and typo behavior.
- Training tools for preparing data, training SAC, packaging models, and visualizing policy behavior.
- Rewritten docs covering runtime usage, architecture, publishing, and RL workflow.
Install
Runtime-only install:
uv add the-bumblebeeInstall RL/model-loading extras:
uv add "the-bumblebee[rl]"Training from a local checkout:
uv sync --group trainRL mouse usage
from bumblebee import Mouse
from bumblebee.rl.policy import SB3MousePolicyPathProvider
provider = SB3MousePolicyPathProvider.from_packaged(deterministic=False)
mouse = Mouse(path_provider=provider)
mouse.move(700, 450)Breaking changes
- Removed the legacy
bumblebee.aipackage and old bundled.pthmodel. - Removed
torchandscipyfrom base runtime dependencies. - RL/training packages are now optional extras instead of default runtime dependencies.
- Old flat RL imports moved to the new package layout:
bumblebee.rl.env→bumblebee.rl.envs.mousebumblebee.rl.gym_env→bumblebee.rl.envs.gymnasiumbumblebee.rl.reward→bumblebee.rl.core.rewardbumblebee.rl.geometry→bumblebee.rl.core.geometrybumblebee.rl.types→bumblebee.rl.core.typesbumblebee.rl.callbacks→bumblebee.rl.training.callbacks
Release assets
| Asset | Purpose |
|---|---|
model.zip |
Same SAC mouse model bundled in the wheel, provided separately for inspection/manual use. |
dataset.npz |
Cleaned mouse demonstration dataset used by the RL workflow. |
No raw tracker JSON files are included.
SHA256
51900ffb4cfd724592235b8e20dfca0858f5b386543ff975650c1cc5d598bd05 model.zip
a2679aa55d7687d55a9b08531d33286e8fc548d9d9f1628a6fdda77a9e8f4e38 dataset.npz