Add HIT policy

[destroy314]

What this does

The Humanoid Imitation Transformer (HIT) is derived from the paper HumanPlus: Humanoid Shadowing and Imitation from Humans (http://arxiv.org/abs/2406.10454), and it is a modification of the ACT model. The key changes include the removal of the VAE and the use of a BERT-style transformer to predict actions and future image embeddings. Although HIT is originally designed for controlling humanoid robots, it can perform better than ACT in the sim ALOHA environment.

How it was tested

• Trained with hit.yaml on lerobot/aloha_sim_insertion_human dataset
• Modified tests/test_available.py by importing new HITPolicy class

How to checkout & try? (for the reviewer)

$ python lerobot/scripts/eval.py -p destroy314/act_aloha_sim_insertion_human eval.n_episodes=100 eval.batch_size=10
{'avg_sum_reward': 261.26, 'avg_max_reward': 2.35, 'pc_success': 15.0, 'eval_s': 439.620165348053, 'eval_ep_s': 4.396201655864716}

$ python lerobot/scripts/eval.py -p destroy314/hit_aloha_sim_insertion_human eval.n_episodes=100 eval.batch_size=10
{'avg_sum_reward': 261.56, 'avg_max_reward': 2.73, 'pc_success': 28.999999999999996, 'eval_s': 439.3340971469879, 'eval_ep_s': 4.3933409762382505}

For training, use

python lerobot/scripts/train.py policy=hit env=aloha env.task=AlohaInsertion-v0 dataset_repo_id=lerobot/aloha_sim_insertion_human

Note that the hyperparameters used for HIT differ from those in the original paper and are instead similar to the ones used in ACT.