Is the model uploaded to Hugging Face the final version?

[mounKim]

I ran an experiment under the conditions below, but the agent only moves for the first 3 seconds, and after that, it keeps outputting the same action repeatedly, like:
OrderedDict([('buttons', 0), ('camera', 220)])
The arguments I used are listed below.

if name == "main":
parser = argparse.ArgumentParser()
parser.add_argument('--workers', type=int, default=1)
parser.add_argument('--split-number', type=int, default=6)
parser.add_argument('--env-config',"-e", type=str, default='smelt/iron_ingot') #vpt/test_vpt
parser.add_argument('--max-frames', type=int, default=400) #vpt/test_vpt
parser.add_argument('--verbos', type=bool, default=False)
parser.add_argument('--checkpoints', type=str, default="CraftJarvis/JarvisVLA-Qwen2-VL-7B")
parser.add_argument('--device',type=str,default="cuda:1")

parser.add_argument('--base-url',type=str, default="http://localhost:8000/v1")
parser.add_argument('--video-main-fold',type=str, default="/workspace/JarvisVLA/logs/")

parser.add_argument('--instruction-type',type=str,default='recipe')
parser.add_argument('--temperature','-t',type=float,default=0.9)
parser.add_argument('--history-num',type=int,default=4)
parser.add_argument('--action-chunk-len',type=int,default=1)

[MuyaoLi-jimo]

I ran an experiment under the conditions below, but the agent only moves for the first 3 seconds, and after that, it keeps outputting the same action repeatedly, like: OrderedDict([('buttons', 0), ('camera', 220)]) The arguments I used are listed below.

if name == "main": parser = argparse.ArgumentParser() parser.add_argument('--workers', type=int, default=1) parser.add_argument('--split-number', type=int, default=6) parser.add_argument('--env-config',"-e", type=str, default='smelt/iron_ingot') #vpt/test_vpt parser.add_argument('--max-frames', type=int, default=400) #vpt/test_vpt parser.add_argument('--verbos', type=bool, default=False) parser.add_argument('--checkpoints', type=str, default="CraftJarvis/JarvisVLA-Qwen2-VL-7B") parser.add_argument('--device',type=str,default="cuda:1")

parser.add_argument('--base-url',type=str, default="http://localhost:8000/v1")
parser.add_argument('--video-main-fold',type=str, default="/workspace/JarvisVLA/logs/")

parser.add_argument('--instruction-type',type=str,default='recipe')
parser.add_argument('--temperature','-t',type=float,default=0.9)
parser.add_argument('--history-num',type=int,default=4)
parser.add_argument('--action-chunk-len',type=int,default=1)

I’m sorry to hear you’re encountering this issue. The current HF Jarvis-VLA is capable of sustained interaction with the environment, but it can occasionally get stuck or fall into repetitive action loops.

You might try running more rollouts to see if the behavior is consistent, or experiment with modifying some sampling parameters, such as --temperature.

If the issue persists, could you share a short video demonstrating the behavior? That would help us better understand and diagnose what’s going on.