Published August 5, 2025 | Version v1

Dynamic Sparsity: Challenging Common Sparsity Assumptions for Learning World Models in Robotic Reinforcement Learning Benchmarks

Authors/Creators

Description

The dataset contains data collected for 1 million time steps across different RL environments in Mujoco Playground using trained-agents with actions sampled from the learnt Gausian distribution. 

Here only the data for the following environments are released as an example - CartpoleBalance, CheetahRun, FingerSpin

Each environment specific folder contains .npz file which has the following keys described in the table below (here num_timesteps=1e6)

Key Definition
states True state values, including initial and terminal states, with shape (num_timesteps x state_dimension)
actions Action values with shape (num_timesteps x action_dimension)
state_jacobians Jacobian matrices representing the derivative of the next true state with respect to the current state, with shape (num_timesteps x state_dimension x state_dimension)
action_jacobians Jacobian matrices representing the derivative of the next true state with respect to the action, with shape (num_timesteps x state_dimension x action_dimension)
obs Observation values with shape (num_timesteps x obs_dim)

rewards

Reward values with shape (num_timesteps x 1)

dones

Done Indicator Vector: A binary vector of shape (num_timesteps, 1) indicating episode terminations. A value of 1 marks a timestep where an episode ends. The corresponding index after this in the states data represents the terminal state. At this terminal index, all keys except for "observation" are set to -1.

total_steps_collected

The total number of timesteps collected

Files

dynamic-sparsity-dataset.zip

Files (3.7 GB)

Name Size
md5:d33bd377c336fa5e00ed68c5eda6695a
3.7 GB Preview Download