File size: 7,086 Bytes
7f97480
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
import os
import json
import argparse
from VLABench.evaluation.evaluator import Evaluator
from VLABench.evaluation.model.policy.openvla import OpenVLA
from VLABench.evaluation.model.policy.base import RandomPolicy
from VLABench.tasks import *
from VLABench.robots import *

os.environ["MUJOCO_GL"]= "egl"

def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--tasks', nargs='+', default=None, help="Specific tasks to run, work when eval-track is None")
    parser.add_argument(
        '--eval-track',
        default=None,
        type=str,
        choices=[
            "track_1_in_distribution",
            "track_sem_all_safe1",
            "track_sem_all_safe2",
            "track_sem_all_unsafe",
            "custom",
        ],
        help="The evaluation track to run"
    )
    parser.add_argument('--n-episode', default=1, type=int, help="The number of episodes to evaluate for a task")
    parser.add_argument('--policy', default="openvla", help="The policy to evaluate")
    parser.add_argument('--model_ckpt', default="/remote-home1/sdzhang/huggingface/openvla-7b", help="The base model checkpoint path")
    parser.add_argument('--lora_ckpt', default="/remote-home1/pjliu/openvla/weights/vlabench/select_fruit+CSv1+lora/", help="The lora checkpoint path")
    parser.add_argument('--save-dir', default="logs", help="The directory to save the evaluation results")
    parser.add_argument('--visulization', action="store_true", default=False, help="Whether to visualize the episodes")
    parser.add_argument('--metrics', nargs='+', default=["success_rate"], choices=["success_rate", "intention_score", "progress_score"], help="The metrics to evaluate")
    parser.add_argument('--host', default="localhost", type=str, help="The host to the remote server")
    parser.add_argument('--port', default=5555, type=int, help="The port to the remote server")
    parser.add_argument('--replanstep', default=5, type=int, help="The step to replan")
    parser.add_argument('--seed-base', default=42, type=int, help="Base seed for episode sampling (seed + episode_id)")
    # VLA-Adapter specific
    parser.add_argument('--vla_ckpt', default=None, type=str, help="Path to VLA-Adapter checkpoint directory")
    parser.add_argument('--vla_num_images', default=2, type=int, help="Number of images used as input (1 or 2)")
    parser.add_argument('--vla_no_proprio', action="store_true", help="Disable proprio input for VLA-Adapter")
    parser.add_argument('--vla_open_loop_steps', default=1, type=int, help="Number of open-loop steps (first executed)")
    parser.add_argument('--vla_skip_center_crop', action="store_true", help="Skip center crop")
    parser.add_argument('--vla_use_minivlm', action="store_true", help="Use mini VLM prompting")
    parser.add_argument('--vla_use_film', action="store_true", help="Enable FiLM in VLA-Adapter vision backbone")
    parser.add_argument('--vla_no_pro_version', action="store_true", help="Disable Pro version head logic")
    parser.add_argument('--vla_camera_index', default=None, type=int, help="Optional camera index override")
    parser.add_argument('--vla_wrist_index', default=None, type=int, help="Optional wrist camera index override")
    parser.add_argument('--action_absolute', action="store_true", help="Treat model outputs as absolute target pose")
    # Nora specific
    parser.add_argument('--nora_camera_index', default=None, type=int, help="Optional camera index override for Nora")
    parser.add_argument('--nora_time_horizon', default=1, type=int, help="Nora action time horizon used by tokenizer")
    args = parser.parse_args()
    return args

def evaluate(args):
    episode_config = None
    if args.eval_track is not None:
        args.save_dir = os.path.join(args.save_dir, args.eval_track)
        with open(os.path.join(os.getenv("VLABENCH_ROOT"), "configs/evaluation/tracks", f"{args.eval_track}.json"), "r") as f:
            episode_config = json.load(f)
            tasks = list(episode_config.keys())
    if args.tasks is not None:
        tasks = args.tasks
    assert isinstance(tasks, list)

    evaluator = Evaluator(
        tasks=tasks,
        n_episodes=args.n_episode,
        episode_config=episode_config,
        max_substeps=1, # repeat step in simulation
        save_dir=args.save_dir,
        visulization=args.visulization,
        metrics=args.metrics,
        seed_base=args.seed_base
    )
    if args.policy.lower() == "openvla":
        policy = OpenVLA(
            model_ckpt=args.model_ckpt,
            lora_ckpt=args.lora_ckpt,
            debug_actions=True,
            norm_config_file=os.path.join(os.getenv("VLABENCH_ROOT"), "configs/model/openvla_config.json"), # TODO: re-compuate the norm state by your own dataset
            action_is_absolute=args.action_absolute

        )
    elif args.policy.lower() == "nora":
        from VLABench.evaluation.model.policy.nora import NoraPolicy

        policy = NoraPolicy(
            model_ckpt=args.model_ckpt,
            device="cuda",
            camera_index=args.nora_camera_index if args.nora_camera_index is not None else 2,
            action_mode="absolute" if args.action_absolute else "delta",
            replan_steps=args.replanstep,
            time_horizon=max(1, args.nora_time_horizon),
        )
    elif args.policy.lower() == "vla_adapter":
        from VLABench.evaluation.model.policy.vla_adapter import VLAAdapterPolicy

        if args.vla_ckpt is None:
            raise ValueError("Please provide --vla_ckpt for VLA-Adapter policy")

        policy = VLAAdapterPolicy(
            checkpoint_path=args.vla_ckpt,
            task_unnorm_key=args.unnorm_key if hasattr(args, "unnorm_key") else "primitive",
            num_images_in_input=args.vla_num_images,
            use_proprio=not args.vla_no_proprio,
            use_film=args.vla_use_film,
            num_open_loop_steps=args.vla_open_loop_steps,
            center_crop=not args.vla_skip_center_crop,
            use_minivlm=args.vla_use_minivlm,
            use_pro_version=not args.vla_no_pro_version,
            device="cuda",
            camera_index=args.vla_camera_index,
            wrist_index=args.vla_wrist_index,
            action_is_absolute=args.action_absolute,
        )
    elif args.policy.lower() == "gr00t":
        from VLABench.evaluation.model.policy.gr00t import Gr00tPolicy
        policy = Gr00tPolicy(host=args.host, port=args.port, replan_steps=args.replanstep)
    elif args.policy.lower() == "openpi":
        from VLABench.evaluation.model.policy.openpi import OpenPiPolicy
        policy = OpenPiPolicy(host=args.host, port=args.port, replan_steps=args.replanstep)
    else:
        policy = RandomPolicy(None)

    result = evaluator.evaluate(policy)
    track = args.eval_track or "custom"
    os.makedirs(os.path.join(args.save_dir, args.policy, track), exist_ok=True)
    with open(os.path.join(args.save_dir, args.policy, track, "evaluation_result.json"), "w") as f:
        json.dump(result, f)

if __name__ == "__main__":
    args = get_args()
    evaluate(args)