Dataset Card for ControlEnvs

Dataset for dynamics learning of 7 classical control tasks based on first-principle models.

Dataset Details

Dataset Description

Seven tasks: Reacher, Pendulum, CartPole, Cart2Pole, Cart3Pole, Acrobot, Acro3bot. Each task has 4 sets of data:

• Training set: 50k sequences of length 64 with random control inputs
• Test set: 20k sequences of length 125 with random control inputs
• Long test set: 10 sequences of length 1000 with random control inputs
• Zero-Control test set: 5k sequences of length 125 with "zero-control" inputs (i.e. no external forces acting on the system)

During dataset generation, the sequences were sampled from longer sequences with random initial conditions chosen to ensure sufficient coverage of the state-space. Coverage of Pendulum, CartPole and Acrobot training sets: TODO

• Curated by: Arne Troch
• Funded by: Research Foundation Flanders (FWO) Grant Number 1SHAH24N
• License: MIT

Dataset Sources [optional]

• Repository: https://github.com/Jianxun-Wang/PIMBRL
• Paper: In progress
• Demo: https://sites.google.com/view/achgn/home

Uses

Direct Use

Intended for learning dynamics of systems subject to external control. Intended for research and benchmarking purposes.

Dataset Structure

Test Data: env/test/data.npz

Long test Data: env/test/data_long.npz

Zero-Control test Data: env/test/data_zero.npz

Train data: env/train/data.npz

Specifications: env/specs.txt

The npz files contain the following keys: ["observations", "actions", "rewards", "states"] The observations and actions are used to learn the dynamics. The rewards can be used when learning a world model aimed at Reinforcement Learning research. The states contain ground-truth information (these are not implemented for all environments! If not implemented, these are the same as the observations.)

Citation

TODO

Dataset Card Contact

Arne Troch: arne.troch@uantwerpen.be