Add Aloha env and ACT policy

WIP Aloha env tests pass

Rendering works (fps look fast tho? TODO action bounding is too wide [-1,1])

Update README

Copy past from act repo

Remove download.py add a WIP for Simxarm

Remove download.py add a WIP for Simxarm

Add act yaml (TODO: try train.py)

Training can runs (TODO: eval)

Add tasks without end_effector that are compatible with dataset, Eval can run (TODO: training and pretrained model)

Add AbstractEnv, Refactor AlohaEnv, Add rendering_hook in env, Minor modifications, (TODO: Refactor Pusht and Simxarm)

poetry lock

fix bug in compute_stats for action normalization

fix more bugs in normalization

fix training

fix import

PushtEnv inheriates AbstractEnv, Improve factory Normalization

Add _make_env to EnvAbstract

Add call_rendering_hooks to pusht env

SimxarmEnv inherites from AbstractEnv (NOT TESTED)

Add aloha tests artifacts + update pusht stats

fix image normalization: before env was in [0,1] but dataset in [0,255], and now both in [0,255]

Small fix on simxarm

Add next to obs

Add top camera to Aloha env (TODO: make it compatible with set of cameras)

Add top camera to Aloha env (TODO: make it compatible with set of cameras)

lerobot/scripts/download.py

@@ -1,22 +0,0 @@
-# TODO(rcadene): obsolete remove
-import os
-import zipfile
-
-import gdown
-
-
-def download():
-    url = "https://drive.google.com/uc?id=1nhxpykGtPDhmQKm-_B8zBSywVRdgeVya"
-    download_path = "data.zip"
-    gdown.download(url, download_path, quiet=False)
-    print("Extracting...")
-    with zipfile.ZipFile(download_path, "r") as zip_f:
-        for member in zip_f.namelist():
-            if member.startswith("data/xarm") and member.endswith(".pkl"):
-                print(member)
-                zip_f.extract(member=member)
-    os.remove(download_path)
-
-
-if __name__ == "__main__":
-    download()

lerobot/scripts/eval.py

@@ -38,27 +38,18 @@ def eval_policy(
     successes = []
     threads = []
     for i in tqdm.tqdm(range(num_episodes)):
-        tensordict = env.reset()
-
         ep_frames = []
-
         if save_video or (return_first_video and i == 0):
 
-            def rendering_callback(env, td=None):
+            def render_frame(env):
                 ep_frames.append(env.render())  # noqa: B023
 
-            # render first frame before rollout
-            rendering_callback(env)
-        else:
-            rendering_callback = None
+            env.register_rendering_hook(render_frame)
 
         with torch.inference_mode():
             rollout = env.rollout(
                 max_steps=max_steps,
                 policy=policy,
-                callback=rendering_callback,
-                auto_reset=False,
-                tensordict=tensordict,
                 auto_cast_to_device=True,
             )
         # print(", ".join([f"{x:.3f}" for x in rollout["next", "reward"][:,0].tolist()]))
@@ -85,6 +76,8 @@ def eval_policy(
             if return_first_video and i == 0:
                 first_video = stacked_frames.transpose(0, 3, 1, 2)
 
+    env.reset_rendering_hooks()
+
     for thread in threads:
         thread.join()
 

lerobot/scripts/train.py

@@ -1,4 +1,5 @@
 import logging
+from pathlib import Path
 
 import hydra
 import numpy as np
@@ -192,6 +193,8 @@ def train(cfg: dict, out_dir=None, job_name=None):
                 num_episodes=cfg.eval_episodes,
                 max_steps=cfg.env.episode_length // cfg.n_action_steps,
                 return_first_video=True,
+                video_dir=Path(out_dir) / "eval",
+                save_video=True,
             )
             log_eval_info(logger, eval_info, step, cfg, offline_buffer, is_offline)
             if cfg.wandb.enable: