mani_skill.trajectory.replay_trajectory

Replay ManiSkill trajectories stored in HDF5 (.h5) format

The replayed trajectory can use different observation modes and control modes.

We support translating actions from certain controllers to a limited number of controllers.

The script is only tested for Panda, and may include some Panda-specific hardcode.

Classes

Args
ReplayResult

Functions

_main(args, use_cpu_backend, env_id, env_kwargs, ...)
_main_helper(x)
main(args)
parse_args([args])
replay_cpu_sim(args, env, ori_env, pbar, episodes, ...)
replay_parallelized_sim(args, env, pbar, episodes, ...)
sanity_check_and_format_seed(episode)	sanity checks the trajectory seed aligns with the episode seed. reformats the reset kwargs seed if missing or formatted wrong

Module Contents

class mani_skill.trajectory.replay_trajectory.Args

allow_failure: bool = False

Whether to include episodes that fail in saved videos and trajectory data based on the environment’s evaluation returned “success” label

count: int | None = None

Number of demonstrations to replay before exiting. By default will replay all demonstrations

discard_timeout: bool = False

Whether to discard episodes that timeout and are truncated (depends on the max_episode_steps parameter of task)

max_retry: int = 0

Maximum number of times to try and replay a trajectory until the task reaches a success state at the end.

num_envs: Annotated[int, tyro.conf.arg(aliases=['-n'])] = 1

Number of environments to run to replay trajectories. With CPU backends typically this is parallelized via python multiprocessing. For parallelized simulation backends like physx_gpu, this is parallelized within a single python process by leveraging the GPU.

obs_mode: Annotated[Optional[str], tyro.conf.arg(aliases=['-o'])] = None

Target observation mode to record in the trajectory. See https://maniskill.readthedocs.io/en/latest/user_guide/concepts/observation.html for a full list of supported observation modes.

record_rewards: bool = False

Whether the replayed trajectory should include rewards

render_mode: str = 'rgb_array'

The render mode used for saving videos. Typically there is also ‘sensors’ and ‘all’ render modes which further render all sensor outputs like cameras.

reward_mode: str | None = None

Specifies the reward type that the env should use. By default it will pick the first supported reward mode. Most environments support ‘sparse’, ‘none’, and some further support ‘normalized_dense’ and ‘dense’ reward modes

save_traj: bool = False

Whether to save trajectories to disk. This will not override the original trajectory file.

save_video: bool = False

Whether to save videos

shader: str | None = None

Change shader used for rendering for all cameras. Default is none meaning it will use whatever was used in the original data collection or the environment default. Can also be ‘rt’ for ray tracing and generating photo-realistic renders. Can also be ‘rt-fast’ for a faster but lower quality ray-traced renderer

sim_backend: Annotated[Optional[str], tyro.conf.arg(aliases=['-b'])] = None

Which simulation backend to use. Can be ‘physx_cpu’, ‘physx_gpu’. If not specified the backend used is the same as the one used to collect the trajectory data.

target_control_mode: Annotated[Optional[str], tyro.conf.arg(aliases=['-c'])] = None

Target control mode to convert the demonstration actions to. Note that not all control modes can be converted to others successfully and not all robots have easy to convert control modes. Currently the Panda robots are the best supported when it comes to control mode conversion. Furthermore control mode conversion is not supported in GPU parallelized environments.

traj_path: str

Path to the trajectory .h5 file to replay

use_env_states: bool = False

Whether to replay by environment states instead of actions. This guarantees that the environment will look exactly the same as the original trajectory at every step.

use_first_env_state: bool = False

Use the first env state in the trajectory to set initial state. This can be useful for trying to replay demonstrations collected in the CPU simulation in the GPU simulation by first starting with the same initial state as GPU simulated tasks will randomize initial states differently despite given the same seed compared to CPU sim.

verbose: bool = False

Whether to print verbose information during trajectory replays

video_fps: int | None = None

The FPS of saved videos. Defaults to the control frequency

vis: bool = False

Whether to visualize the trajectory replay via the GUI.

class mani_skill.trajectory.replay_trajectory.ReplayResult

num_replays: int

successful_replays: int

mani_skill.trajectory.replay_trajectory._main(args, use_cpu_backend, env_id, env_kwargs, ori_env_kwargs, record_episode_kwargs, proc_id=0, num_procs=1)

Parameters:

• args (Args) –

• proc_id (int) –

mani_skill.trajectory.replay_trajectory._main_helper(x)

mani_skill.trajectory.replay_trajectory.main(args)

Parameters:

args (Args) –

mani_skill.trajectory.replay_trajectory.parse_args(args=None)

mani_skill.trajectory.replay_trajectory.replay_cpu_sim(args, env, ori_env, pbar, episodes, trajectories)

Parameters:

• args (Args) –

• env (mani_skill.utils.wrappers.record.RecordEpisode) –

mani_skill.trajectory.replay_trajectory.replay_parallelized_sim(args, env, pbar, episodes, trajectories)

Parameters:

• args (Args) –

• env (mani_skill.utils.wrappers.record.RecordEpisode) –

mani_skill.trajectory.replay_trajectory.sanity_check_and_format_seed(episode)

sanity checks the trajectory seed aligns with the episode seed. reformats the reset kwargs seed if missing or formatted wrong