README.md

Bin Pick and Place Tasks (Debug)

We want to eventually learn two primitives: pick up a block and place it into a bin, and pick up a block from the bin and place it somewhere else. Not all of these tasks are learnable directly.

Update: The culprit was a missing reset in _reset_sim. because of this, the arm wonders around, drifting through a large configuration space.

Now, all environments are working.

Name	Observation Spec	Goal Init/Comment	Status
Bin-no-bin-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	env uses bin_null.xml, other wise sam	✅ 99%
Bin-no-bin-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	env is the same as fetch:PickPlace-v0	✅ 99%
Bin-no-init-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	same except bin in model xml	✅ 99%
Bin-aside-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	same except bin in model	✅ 99%
Bin-aside-pos-v0	achieved_goal: (3,) desired_goal: (3,) observation: (28,)	plus 3-dim position in obs	✅ 99%

Details of Each Task

Initial	After Reset	After Reset

render_initial('fetch:Bin-no-bin-v0', row)
render_video('fetch:Bin-no-bin-v0', 15, row)

Initial	After Reset	After Reset

render_initial('fetch:Bin-pp-xml-v0', row)
render_video('fetch:Bin-pp-xml-v0', 15, row)

Initial	After Reset	After Reset

render_initial('fetch:Bin-no-init-v0', row)
render_video('fetch:Bin-no-init-v0', 15, row)

Initial	After Reset	After Reset

render_initial('fetch:Bin-aside-v0', row)
render_video('fetch:Bin-aside-v0', 15, row)

Initial	After Reset	After Reset

render_initial('fetch:Bin-aside-pos-v0', row)
render_video('fetch:Bin-aside-pos-v0', 15, row)

README.py

import gym
import numpy as np
from cmx import doc

scale = 3
src_prefix = "206a41c28e9a0c6953c9ca0f68a0210f"


def get_obs_spec(env_id):
    env = gym.make(env_id)
    buffer = []
    for k, v in env.observation_space.spaces.items():
        buffer += [f"{k}: {v.shape}"]
    return "<br>".join(buffer)


def render_initial(env_id, doc):
    env = gym.make(env_id)
    img = env.render('rgb_array', width=200 * scale, height=120 * scale)
    doc.figure(img, src=f"{src_prefix}/{env_id}_init.png?ts={doc.now('%f')}", title="Initial")

    frames = []
    for i in range(10):
        env.reset()
        frames.append(env.render('rgb_array', width=100 * scale, height=120 * scale))
    env_id = env_id.split(':')[-1]
    doc.figure(np.array(frames).min(axis=0), src=f"{src_prefix}/{env_id}_reset.png?ts={doc.now('%f')}",
               title="After Reset")


def render_video(env_id, n, doc):
    env = gym.make(env_id)
    frames = []
    for ep in range(n):
        env.reset()
        for i in range(10):
            act = env.action_space.sample()
            ts = env.step(act)
            frames.append(env.render('rgb_array', width=100 * scale, height=120 * scale))

    env_id = env_id.split(':')[-1]
    doc.video(np.array(frames), src=f"{src_prefix}/{env_id}.gif?ts={doc.now('%f')}",
              title="After Reset")


if __name__ == '__main__':
    doc.config("README.md")
    doc @ f"""
# Bin Pick and Place Tasks (Debug)

We want to eventually learn two primitives: pick up a block and 
place it into a bin, and pick up a block from the bin and place
it somewhere else. Not all of these tasks are learnable directly.

> **Update**: The culprit was a missing reset in `_reset_sim`. 
> because of this, the arm wonders around, drifting through
> a large configuration space.

Name                 | Observation Spec                         | Goal Init/Comment                        | Status
-----------------    | ----------------                         | -------                                  | ------
**Bin-no-bin-v0**    | {get_obs_spec('fetch:Bin-no-bin-v0')}    | env uses `bin_null.xml`, other wise sam  | ✅ 99%
**Bin-no-bin-v0**    | {get_obs_spec('fetch:Bin-pp-xml-v0')}    | env is the same as `fetch:PickPlace-v0`  | ✅ 99%
**Bin-no-init-v0**   | {get_obs_spec('fetch:Bin-no-init-v0')}   | same except bin in model xml             | ✅ 99%
**Bin-aside-v0**     | {get_obs_spec('fetch:Bin-aside-v0')}     | same except bin in model                 | ✅ 99%
**Bin-aside-pos-v0** | {get_obs_spec('fetch:Bin-aside-pos-v0')} | plus 3-dim position in obs               | ✅ 99%

## Details of Each Task
"""
    table = doc.table()
    with doc, table.figure_row() as row:
        render_initial('fetch:Bin-no-bin-v0', row)
        render_video('fetch:Bin-no-bin-v0', 15, row)

    table = doc.table()
    with doc, table.figure_row() as row:
        render_initial('fetch:Bin-pp-xml-v0', row)
        render_video('fetch:Bin-pp-xml-v0', 15, row)

    table = doc.table()
    with doc, table.figure_row() as row:
        render_initial('fetch:Bin-no-init-v0', row)
        render_video('fetch:Bin-no-init-v0', 15, row)

    table = doc.table()
    with doc, table.figure_row() as row:
        render_initial('fetch:Bin-aside-v0', row)
        render_video('fetch:Bin-aside-v0', 15, row)

    table = doc.table()
    with doc, table.figure_row() as row:
        render_initial('fetch:Bin-aside-pos-v0', row)
        render_video('fetch:Bin-aside-pos-v0', 15, row)

    doc.flush()

Bin-no-bin-v0.gif

Bin-no-bin-v0_init.png

Bin-no-bin-v0_reset.png

train.py

from rl import main, Args
from params_proto.neo_hyper import Sweep
from ml_logger import logger

if __name__ == '__main__':
    import jaynes
    from lp_analysis import instr

    env_names = [
        # Gym originals
        'fetch:Slide-v0',
        'fetch:PickPlace-v0',
        # Bin debug task set
        'fetch:Bin-no-bin-v0',
        'fetch:Bin-pp-xml-v0',
        'fetch:Bin-no-init-v0',
        'fetch:Bin-aside-v0',
        'fetch:Bin-aside-pos-v0',
        # # Bin debug task sets
        # 'fetch:Bin-fixed-v0',
        # 'fetch:Bin-fixed-hide-v0',
        # 'fetch:Bin-fixed-pos-v0',
    ]

    with Sweep(Args) as sweep:

        Args.gamma = 0.99
        Args.clip_inputs = True
        Args.normalize_inputs = True
        Args.n_workers = 12
        Args.n_epochs = 500

        with sweep.product:
            with sweep.zip:
                Args.env_name = env_names

            Args.seed = [100, 200, 300, 400, 500, 600]

    # jaynes.config('local' if Args.debug else "gpu-vaughan", launch=dict(timeout=10000))
    jaynes.config('local' if Args.debug else "cpu-mars")
    # jaynes.config('local')
    for i, deps in sweep.items():
        thunk = instr(main, deps, _job_postfix=f"{Args.env_name.split(':')[-1]}")
        logger.log_text("""
            keys:
            - run.status
            - Args.env_name
            - host.hostname
            charts:
            - yKey: test/success
              xKey: epoch
            - yKey: test/success
              xKey: timesteps
            - yKey: train/success/mean
              xKey: time
            """, ".charts.yml", dedent=True)
        jaynes.run(thunk)