OCBench is a controllable robotic manipulation benchmark for studying behavioral cloning (BC) and reinforcement learning (RL). OCBench provides GPU-accelerated tasks and scripted policies that mimic the properties of human demonstrations. This enables controlled studies of BC and RL algorithms with virtually unlimited data.

OCBench can be easily installed via PyPI:

pip install "ocbench[mjwarp,train]"

The additional mjwarp and train dependencies are required for the GPU-accelerated (MJWarp) environments and reference implementations, respectively. If you only want to use the CPU environments, you can install OCBench without the optional dependencies:

pip install ocbench

OCBench environments follow the Gymnasium interface. Here's an example of instantiating and running a CPU environment:

import ocbench

# Make an environment.
env = ocbench.make('block-cpu-single-task1-v0')
ob, info = env.reset()

# Run an episode.
ret = 0.0
done = False
while not done:
    action = env.action_space.sample()  # Replace this with your agent's action.
    ob, reward, terminated, truncated, info = env.step(action)
    ret += reward
    done = terminated or truncated

print('Return:', ret)
env.close()

OCBench defaults to GPU-accelerated (MJWarp) environments. Here's an example of running 1024 parallel environments:

import numpy as np
import ocbench

# Make 1024 parallel environments.
env_name = 'block-single-task1-v0'
env = ocbench.make(env_name, nworld=1024)
_, _, _, max_episode_steps = ocbench.parse_env_spec(env_name)
ob, info = env.reset()

# Run 1024 episodes in parallel.
episode_returns = np.zeros(env.nworld)
done = np.zeros(env.nworld, dtype=bool)
for step in range(max_episode_steps):
    action = env.action_space.sample()  # Shape: (1024, 7). Replace with your agent's actions.
    action[done] = 0.0
    ob, reward, terminated, truncated, info = env.step(action)
    episode_returns[~done] += reward[~done]
    done |= terminated | truncated
    if done.all():
        break

print('Mean return:', episode_returns.mean())
env.close()

OCBench provides 28 manipulation tasks across 5 types of environments. See the paper for more details.

OCBench also provides several variants of the environments:

• lite: The lite variant has much shorter horizons and simplified scripted policies. This can be useful for faster prototyping.
• cpu: The cpu variant uses CPU-based MuJoCo instead of GPU-based MJWarp.
• visual: The visual variant provides pixel-based observations instead of state-based observations.

OCBench datasets are hosted on Hugging Face. The datasets will be automatically downloaded to ~/.cache/ocbench/datasets by default. You can also specify a dataset root directory or path by setting dataset_path or dataset_root.

For lite, cpu, and visual variants, we only provide datasets for selected tasks. In particular, we do not provide datasets for full-horizon cpu environments. However, you can still generate datasets for these variants using the provided data-generation script.

Here's an example of how to load an environment and its datasets:

import ocbench

# Make an environment and load datasets.
env_name = 'block-single-task1-v0'
env, train_dataset, val_dataset = ocbench.make_env_and_datasets(
    env_name,  # Environment name.
    dataset_root='~/.cache/ocbench/datasets',  # Directory to save datasets (optional).
    nworld=1024,  # Number of parallel environments (optional).
)

# Example: Episode 1 has 3 steps and terminates early;
# episode 2 has 4 steps and reaches the time limit.
#
#                  |<-episode 1->|  |<---episode 2--->|
# ---------------------------------------------------------------
# 'observations': [s0, s1, s2, s3,  s0, s1, s2, s3, s4,  ...]
# 'actions'     : [a0, a1, a2,      a0, a1, a2, a3,      ...]
# 'rewards'     : [r0, r1, r2,      r0, r1, r2, r3,      ...]
# 'terminals'   : [ 0,  0,  1,       0,  0,  0,  1,      ...]
# 'masks'       : [ 1,  1,  0,       1,  1,  1,  1,      ...]
#
# Each episode includes its final observation, but no extra action, so
# the 'observations' and 'actions' arrays can have different lengths.
# For observation_interval > 1 (i.e., sparse datasets), observations are stored
# at steps 0, interval, 2 * interval, ..., plus the final observation.

We provide scripted policies and a data-generation script in data_gen_scripts/generate_ocbench.py. We provide the exact commands for the datasets used in the paper in commands.sh.

OCBench also supports streaming (on-the-fly) data generation for MJWarp environments to train an agent with virtually unlimited data. Set --streaming_data=1 and related flags in main.py to use this feature.

We provide reference implementations of several BC and RL algorithms in the impls directory.

After cloning this repository, run the following from its root directory:

cd impls
pip install "ocbench[mjwarp,train]"

By default, this uses the PyPI version of OCBench. To use a local version of OCBench (e.g., for training on modified environments), run the following instead from the repository root:

pip install -e ".[mjwarp,train]"

• BC (bc.py): Deterministic behavioral cloning.
• FBC (fbc.py): Flow behavioral cloning.
• FSQBC (fsqbc.py): FSQ behavioral cloning.
• History-conditioned FBC (history_fbc.py): History-conditioned flow behavioral cloning.
• IFQL (ifql.py): Implicit flow Q-learning (the flow variant of implicit diffusion Q-learning).
• FQL (fql.py): Flow Q-learning.
• DSRL (dsrl.py): Diffusion steering via reinforcement learning (DSRL-NA).
• Advantage conditioning (binary_advc.py): Binary advantage-conditioned flow behavioral cloning.

We provide commands to reproduce the benchmark tables in the paper in commands.sh.

This codebase is inspired by or partly uses code from the following repositories:

• OGBench for the general infrastructure.
• MuJoCo and MJWarp for simulation.
• MuJoCo Menagerie for the robot descriptions (Universal Robots UR5e and Robotiq 2F-85).
• Meta-World for the objects (drawer, window, and button) in the manipulation environments.

@article{park2026behavioral,
  title={{Behavioral Cloning Mystery}},
  author={Park, Seohong and Levine, Sergey},
  journal={ArXiv},
  year={2026},
  volume={abs/2610.07056},
}