Add real-world support for ACT on Aloha/Aloha2 (#228)

Co-authored-by: Alexander Soare <alexander.soare159@gmail.com>
这个提交包含在:
Remi
2024-05-31 15:31:02 +02:00
提交者 GitHub
父节点 504d2aaf48
当前提交 d585c73f9f
共有 22 个文件被更改,包括 525 次插入 和 140 次删除

查看文件

@ -75,15 +75,16 @@ def get_policy_stats(env_name, policy_name, extra_overrides):
# HACK: We reload a batch with no delta_timestamps as `select_action` won't expect a timestamps dimension
dataset.delta_timestamps = None
batch = next(iter(dataloader))
obs = {
k: batch[k]
for k in batch
if k in ["observation.image", "observation.images.top", "observation.state"]
}
obs = {}
for k in batch:
if k.startswith("observation"):
obs[k] = batch[k]
if "n_action_steps" in cfg.policy:
actions_queue = cfg.policy.n_action_steps
else:
actions_queue = cfg.policy.n_action_repeats
actions_queue = (
cfg.policy.n_action_steps if "n_action_steps" in cfg.policy else cfg.policy.n_action_repeats
)
actions = {str(i): policy.select_action(obs).contiguous() for i in range(actions_queue)}
return output_dict, grad_stats, param_stats, actions
@ -114,6 +115,8 @@ if __name__ == "__main__":
["policy.n_action_steps=8", "policy.num_inference_steps=10", "policy.down_dims=[128, 256, 512]"],
),
("aloha", "act", ["policy.n_action_steps=10"]),
("dora_aloha_real", "act_real", ["policy.n_action_steps=10"]),
("dora_aloha_real", "act_real_no_state", ["policy.n_action_steps=10"]),
]
for env, policy, extra_overrides in env_policies:
save_policy_to_safetensors("tests/data/save_policy_to_safetensors", env, policy, extra_overrides)