PushT Dataset

Video + action data from the gym-pusht environment. Three splits:

Split	Episodes	Avg steps/ep	Hours	Description
`smooth/`	~38,900	300	~324 hrs	Random smooth movement (Ornstein-Uhlenbeck process)
`goal/`	~8,700	298	~73 hrs	Heuristic goal-directed policy (keypoint matching)
`expert/`	~21,800	228	~138 hrs	Pretrained diffusion policy, ~74% success rate

File format

Each .npz file contains multiple episodes. Load with:

import numpy as np
data = np.load("smooth/smooth_0000_00.npz", allow_pickle=True)
n = int(data["num_trajectories"])  # number of episodes in this file
for i in range(n):
    frames  = data[f"frames_{i}"]   # (T+1, 96, 96, 3) uint8 — RGB pixel observations
    actions = data[f"actions_{i}"]  # (T, 2)             float32 — agent target position [x, y] in [0, 512]
    rewards = data[f"rewards_{i}"]  # (T,)               float32 — coverage ratio, 1.0 = solved
    policy  = str(data[f"policy_{i}"])  # "smooth", "goal", or "expert"

frames has one more entry than actions (initial frame before first action)
frames[t] is the observation before actions[t] is taken
frames[t+1] is the observation after actions[t] is taken
The environment runs at 10 Hz (0.1s per step)
An episode is "solved" when rewards[t] >= 0.95 (the T-block covers >95% of the goal)