search

MIT-TESSE / goseek-challenge

Instructions for competing in the GOSEEK challenge at ICRA 2020
67 stars 16 forks source link

Simulator Stop(version 0.1.3) #7

Closed ZHMA1996 closed 4 years ago

ZHMA1996 commented 4 years ago

Is is possible that the simulator would suddenly stop ?

when testing the simulator with an agent taking random action, the simulator just go forward several

steps and just stop. And no similar phenomenon was observed in version 0.1.0.

ZHMA1996 commented 4 years ago

Especially when the agent take the action 3

ZHMA1996 commented 4 years ago

It seems the program stops in the function advance_game_time, could anyone explain in detail and how to solve it.

ZacRavichandran commented 4 years ago

Hi,

Did this happen when running:

python eval.py --agent-config baselines/config/random-agent.yaml --episode-config config/check-ground-truth.yaml

or are there other cases where you've observed that behavior?

ZHMA1996 commented 4 years ago

Hi

when running the command you provided, this behavior didn't happen.

ZacRavichandran commented 4 years ago

Good to know. Is there a command we can use to replicate the issue?

ZHMA1996 commented 4 years ago

When testing my own agent, the behavior happened. I tried to figure out why and use a for loop to

iterate over every step and found that the simulator would stop when fetching fruits. In this setting, I

created a Vector Env using the library, stable-baselines. After I commented out the function

advance_game_time in the source code and then the simulator could work again.

ZacRavichandran commented 4 years ago

Thanks for the description. If you're still having issues, could you try setting the step_rate argument in GoseekFullPerception to 20 (doesn't have to be 20 specifically, but in that range would be good).

In the baseline notebook, this would look like:

env = GoSeekFullPerception(
                str(filename),
                step_rate=20,
                network_config=get_network_config(worker_id=rank),
                n_targets=n_targets,
                episode_length=episode_length,
                scene_id=scene_id[rank],
                target_found_reward=target_found_reward,
            )
ZHMA1996 commented 4 years ago

Thanks for your reply. This problem could be solved after setting the step_rate argument !

ZacRavichandran commented 4 years ago

Good to know! I'll update the baseline notebook to suggest that!

pushkalkatara commented 4 years ago

Faced same issue in v0.1.4. Adding step_rate=20 solved the issue.