openai / neural-mmo

Code for the paper "Neural MMO: A Massively Multiagent Game Environment for Training and Evaluating Intelligent Agents"

Home Page:https://openai.com/blog/neural-mmo/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Agent pass the stone

wwxFromTju opened this issue · comments

Hi, when i try to render the env, I find that agent pass the stone.

2019-03-05 6 42 40
2019-03-05 6 42 14

Odd. I haven't observed that. Is this consistent, or on loadup in the middle of a lagspike? Want to make sure it's just not "teleporting" through due to skipping a few frames before looking deeper

I first run: python Forge.py --render
Then use chrome to enter: http://localhost:8080/forge/embyr/
Log info is below:

reated a server WebSocket connection request: {"peer": "tcp4:127.0.0.1:53419", "headers": {"host": "127.0.0.1:8080", "connection": "Upgrade", "pragma": "no -cache", "cache-control": "no-cache", "upgrade": "websocket", "origin": "http://localhost:8080", "sec-websocket-version": "13", "user-agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_14_2) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/72.0.3626.119 Safari/537.36", "accept-encoding": "gzip, deflate, br ", "accept-language": "zh-CN,zh;q=0.9", "sec-websocket-key": "Qky+LY2S8jvF72mWPdgIyw==", "sec-websocket-extensions": "permessage-deflate; client_max_window_bits "}, "host": "127.0.0.1", "path": "/ws", "params": {}, "version": 13, "origin": "http://localhost:8080", " Protocols:: [], "extensions": [["permessage-deflate", {"client_max_window_bits": [true]}]]} Opened connection to server

Then I use wsad or up, down, left, right for fun, than find agent pass the stone.
For example below:
2019-03-06 7 14 18
2019-03-06 7 12 37

I didn't just use zoom in and out to see the map.

I observed that the Forest near the agent is occasionally collected. I think the agent may drift when it is a render.

Ah okay, good. I was going to say, I'm fairly positive it's correct in the underlying environment. Probably at worse a 1 tick desync or off by 1 error in the renderer. Traveling ATM, but I'll look into it over the next few days. If you happen to find it before me, feel free to submit a PR to the client repo

I've observed a similar issue. It seems to happen consistently when I render the env.
Screen Shot 2019-03-10 at 10 44 34 PM

Alright, well aware of this now. Again, probably an off by 1 in the renderer. Will look into this as soon as I get a chance.

I'm only able to reproduce this during intentionally induced lag spikes (expected behavior -- if you miss a few ticks, agents move in a direct line to their final position). Tried wasd, arrow keys, mouse control -- nothing seems to cause this behavior after about 15-20 minutes of testing. I'm getting 50 FPS on Firefox and double that on Chrome. I'm thinking there might be some weird lag induced desync on lower end systems. What frame rates are you getting after the environment has warmed up and there are agents all over the map?

I added manual control for one of the agents and found that the position in the view is off by one in x and off by one in y as compared to the model. Perhaps there is an off-by-one error? The game would pause after every tick and await user input, so it may not be frame rate related.

yep, I fix it by: #22

The same happens with lava:
in_lava
What is not obvious from a still image is that some of the agents will then die but some others will come back to land safe and sound.

Just FYI -- have not forgotten about this, just got caught up with the awesome DotA result and working on the first major environment patch. You're probably right that I'm using a different code version. I'll likely bundle this fix with the patch once I get to it

Well, something is really screwy here. I just made sure I was on the same code version (redid set up from scratch), tested both Chrome and Firefox, and I still can't reproduce the issue. If you are not getting this error, please post to let me know -- I'd like to know if it's on my end (despite testing on multiple machines) or if it's setup dependent.

Update: spent a long time trying to reproduce this, have tested on multiple machines, firefox + chrome ... it's gone in 1.1. Otherwise reopen.