decisionforce / pgdrive Goto Github PK

The pre-generated maps are identical to previous version, verified in Dec 19, 2020.

Link to #14

The allowable display:

• "pandagl": Error
• "pandagles", Not detected
• "pandagles2", Not detected
• "pandadx9", Not detected
• "p3tinydisplay", Bug-free ✅
• pandadx9, Not detected
• pandadx8
• tinydisplay
• pdtinygl

Ref:

https://docs.panda3d.org/1.10/python/programming/configuration/configuring-panda3d

https://docs.panda3d.org/1.10/python/programming/texturing/texture-compression?highlight=pandagles

If so, we can leverage more workers in RGB training!

To prevent new PR from introducing new bug, we need some unit tests to do CI

So that we can smooth the edge between sky and grassland! See: https://docs.panda3d.org/1.10/python/programming/render-attributes/fog

If serious bug happen, the vehicle is flipped, it will not reset the state.

Run test_random_traffic.py in fix-random branch:

After a reset:

You can see that the color is changing, but the reported location and car type is not changed.

I am wondering why this happen?

We should make "random_traffic = False" mode completely reproducible.

SkyBox

• Tear on OpenGL (MacOS)
• Tear on tinydisplay (All Platform)
• Tear on OpenGLES (Headless Machine)

Other Vehicle Model

• Load incorrectly on Tinydisplay (All Platform)
• Load incorrectly on OpenGLES (All Platform)
• Load incorrectly on OpenGL (MacOS with OpenGL 3 2)
• Too Big, making repo big & making training resources exhausted

In that case, current save and load maps mechanism will raise error.

This is dangerous for map-generation.

1. Layout error

2. Possibly size error?

Is ego vehicle has correct size compared to traffic vehicle?

I suggest stop developing this thing before release of v0.1.0

This happen in python -m pgdrive.examples.enjoy

Is this correct?

It's seldom to have multiple pg_world.

We should destory the vehicle in either destory of __del__

• Repo name change to pgdrive
• System name (in setup.py's description and Readme, logo) change to PGDrive
• python package change to pgdrive (everything in code and setup.py's name)

I need to train again for a super agent.

• Train
• Extract weights
• Remove historical agent in all places (bfs)
• Build numpy agent
• Upload numpy agent

Professor says when using enjoy script the FPS slows down gradually.

• mode != "human" return current image
• use igloop to do render instead of calling renderFrame() out of the loop
• move main camera moving and other update into taskMgr to do it automatically

Quanyi has made a ResetEnv before, with the capacity to use keyboard stroke to reset the environment. I think that's good to integrate such functionalities into the default environment when using manual controlling.

We can also think of other useful commands that we wish to tell the environment, for example the maximum FPS.

The primal goal:

Use the trajectory from lidar-specified environment, to generate a new RGB-image observation trajectory.

This is the preliminary functionality, before we wish to use lidar-based agent to generate image-based trajectory for imitation learning.

To Do List:

• Can single vehicle state be perfectly restored?
• Can opponent vehicle state be perfectly restored?
• Can the image observation be perfectly restored?

We should not assume user's gym version!

When you set camera_height to 800 to see the overview of the whole scenes, you will find that during reset, the environment is reverse for a moment.

Reproduction:

1. Arbitrary multi-map environment,
2. camera_height to 800
3. Keep resetting the environment.

Before reset (last environment):

First moment when resetting:

Third moment when resetting:

The agent driving in the scene:

To prevent PR from introducing new bugs, we need a zombie car to help us check.

Press once: force to 60
Press twice: force to realtime

The episode reward stop growing and achieve similar episode reward as in 3 blocks environments.

The training success rate is quite low for 7 blocks even the reward of 7 blocks and 3 blocks are similar.

@lqy0057

This kind of message is distracting.

    env = PGDriveEnv({
        "environment_num": 10000,
        "traffic_density": 0.2,
        "map_config": {
            Map.GENERATE_METHOD: MapGenerateMethod.BIG_BLOCK_NUM,
            Map.GENERATE_PARA: 7,
        }
    })

    env = PGDriveEnv({
        "environment_num": 10000,
        "traffic_density": 0.5,
        "map_config": {
            Map.GENERATE_METHOD: MapGenerateMethod.BIG_BLOCK_NUM,
            Map.GENERATE_PARA: 7,
        }
    })

@lqy0057 Please take a look on: add-helper branch, pgdrive/tests/generalization_env_test/test_vehicle_num.py

This issue is for reference only. We believe this bug is fixed after #86 and #146 .

If any user find this issue again, please report to us, thanks!

Keywords: memory leak, render, rendering, pipeline, model, texture, compress, compression, break, broken, gles, Mac, platform, onscreen, window, pgdrive v0.1.0

• 1. Final update of README, with carefully polished languague

• 2. Final update of Documents

• 3. Upload a new expert agent

• 4. Check all test scripts is working (remove failed ones)

• 5. Check all examples scripts is working (three local-run examples, colab script should wait the publication of repo.)

Current package size is too large for installation!
@lqy0057

As you can see in the above figure.

I think we should go through the code and see how many task are there in the task manager. Maybe we can query the taskmgr for the full list and close all tasks together.

You should press twice "1" in order to see the box debug result.

Including:

• 1. Scalar observation environment go through
• 2. Image-based observation go through (colab dependencies installation is also important)
• 3. How to specify the config (change density, blocks nums and so on)
• 4. The BIG algorithm bird-view demonstrations (a video or gif would be useful)
• 5. Use my super agent to run the environment (can be applied to all 4 examples)

I don't think this is normal. We should have identical behavior in all platform.

We notice extremely large step reward.

After fixing the velocity #125 #129 , no extremely small step reward found:

Let's take a look:

        current_lane = self.vehicle.lane
        long_last, _ = current_lane.local_coordinates(self.vehicle.last_position)
        long_now, lateral_now = current_lane.local_coordinates(self.vehicle.position)

        # Suspicious!
        reward = 0.0
        lateral_factor = 1 - 2 * abs(lateral_now) / self.current_map.lane_width
        reward += self.config["driving_reward"] * (long_now - long_last) * lateral_factor

        # Suspicious!
        steering_change = abs(self.vehicle.last_current_action[0][0] - self.vehicle.last_current_action[1][0])
        steering_penalty = self.config["steering_penalty"] * steering_change * self.vehicle.speed / 20
        reward -= steering_penalty

        # Zero!
        acceleration_penalty = self.config["acceleration_penalty"] * ((action[1])**2)
        reward -= acceleration_penalty

        # Zero!
        low_speed_penalty = 0
        if self.vehicle.speed < 1:
            low_speed_penalty = self.config["low_speed_penalty"]  # encourage car
        reward -= low_speed_penalty

        # Zero!
        reward -= self.config["general_penalty"]

        # Suspicious!
        reward += self.config["speed_reward"] * (self.vehicle.speed / self.vehicle.max_speed)

The driving status label and white line framework are incorrect, when using different resolution

Can we suppress pygame message?

2020-12-20 09:58:25.610 python[17378:331264] IMKClient Stall detected, *please Report* your user scenario attaching a spindump (or sysdiagnose) that captures the problem - (imkxpc_bundleIdentifierWithReply:) block performed very slowly (5.89 secs).

Is this msg what we expected?

There are still references of ImageBuffer in pgworld class and python will not release these useless things. It's ok when we only create vehicle for once, but creating and destroying vehicle for several times will cause critical bugs due to memory leak (segment fault or other problem)

(Feel free for the agents training with lidar. This bug will not affect them

The following config can be changed and take effect in next episode:

• mannual control
• traffic density
• traffic mode
• image source (buffer size should be same when change)
• decision repeat
• the whole reward scheme
• map config, when fall back to use BIG online (need special process,)

Currently, the map seed and traffic seed are coupled in PGDriveEnv.
It would be hard to re-implement works based on Carla. Since works on carla use fixed maps with changing dynamic factors (ego car's spawn place / target place, other vehicle's spawn place, ...)
Also, I am training IL algorithm on 1-env setting and want to collect some expert data. I utilized pgdrive.examples.expert but to find out that the collected trajectories were almost the same!

So, shall we provide an environment with decoupling map seed and traffic seed? Like:
env = PGDriveDecouplingEnv(dict=(environment_num=1, traffic_num=100))

The outskirt of the front-cam in visualization console is mis aligned.

A top-down view like highway-env to support more research topics. Besides, Carla and other simulators also provide a simple top down view render and multi-layer image observation.

I will add navigation line soon

depth相机给出来的画面挺诡异的

总之我没有要找到一个文件能够合理的展示depth相机的黑白画面（如同你上次展示的那样）

我对他的功能还是存疑的，如果他只是把地板和天空给过滤掉的话，那和原来的应该有的功能差别还是太大了

We should:

1. Use "highway" or other string to represent the "highway_mode", instead of spliting self.pg_world.mode and self.pg_world.pg_config["highway_render"].
2. Stop querying self.pg_world.pg_config["highway_render"] in all places inside pg_world.

Therefore, we will have four strings to represent render mode in the internal PGWorld now:

1. none
2. onscreen
3. offscreen
4. highway (or other name, I am not sure, like birdview or topdown)