sim/save_policy_rollout_gt_sim.py

"""
Generate data based on the learned policy and physics simulator - mujoco
"""

from sim.simulator import RobomimicSimulator
from sim.policy import DiffusionPolicy
from diffusion_policy.util.pytorch_util import dict_apply

import h5py
import tqdm
import numpy as np
import torch
import cv2
import imageio

MAX_STEPS = 100
RES = 84    # for DP input

if __name__ == '__main__':
    
    env = RobomimicSimulator(env_name='lift')
    policy = DiffusionPolicy('data/dp_ckpt/dp_lift_sr0.70.ckpt')
    
    n_obs_steps = policy.n_obs_steps

    demos = dict()
    for trial in range(200):
        # reset all
        image = env.reset()

        this_demo = {
            "images": [],
            "actions": []
        }
        
        latest_obs_dict = {"agentview_image": cv2.resize(image, (RES, RES)).transpose(2, 0, 1)}
        obs_dict_buf = dict_apply(latest_obs_dict, lambda x: x[np.newaxis].repeat(n_obs_steps, axis=0))
        
        done = False
        pbar = tqdm.tqdm(total=MAX_STEPS)
        while not done and pbar.n < MAX_STEPS:

            # get latest obs
            latest_obs_dict = {"agentview_image": cv2.resize(image, (RES, RES)).transpose(2, 0, 1)}
            obs_dict_buf = dict_apply(obs_dict_buf, lambda x: np.roll(x, -1, axis=0))
            for k, v in latest_obs_dict.items():
                obs_dict_buf[k][-1] = v

            # rollout
            traj = policy.generate_action(dict_apply(
                obs_dict_buf,
                lambda x: torch.from_numpy(x).to(
                    device=policy.device, dtype=policy.dtype
                    ).unsqueeze(0)
            ))['action'].squeeze(0).detach().cpu().numpy()

            # step the simulator
            for action in traj:
                this_demo["images"].append(image)
                this_demo["actions"].append(action)

                result = env.step(action)
                done = done or result['done']
                image = result['pred_next_frame']

                pbar.update(1)

        this_demo = dict_apply(this_demo, lambda x: np.array(x))
        demos[f"demo_{trial}"] = this_demo


    demos = {"data": demos}
    with h5py.File('data/my_robomimic_dataset.hdf5', 'w') as f:
        # save demos
        """
        demos = {
            "data": {
                "demo_0": {
                    "images": np.array([...]),
                    "actions": np.array([...])
                },
                "demo_1": {
                    "images": np.array([...]),
                    "actions": np.array([...])
                },
                ...
            }
        }
        """
        data = f.create_group("data")
        for demo_name, demo_data in demos["data"].items():
            demo = data.create_group(demo_name)
            for key, value in demo_data.items():
                demo.create_dataset(key, data=value)