4. Logger

Overview

RoboHive provides a custom logger to facilitate working with grouped data. Logger consists of a name for the logs, group for organizing the data, and dataset for storing data. Attached below is logger's general schema -

name: {
    group{g1: dataset{k1:v1}, dataset{k2:v2}, ...}
    group{g2: dataset{kx:vx}, dataset{ky:vy}, ...}
    ...
}

Let's understand RoboHive's Logger details using its most common use case -- i.e. recording Robot trajectories.

Robot Trajectories (Rollouts)

Robot trajectories, known as rollouts, are represented in RoboHive as time aligned Tuples. Note that for a trajectory of horizon h we will have h+1 tuples to account for the initial and the final states. We record the last actions as NaN

<t(0)   : s(0),     a(0)>
<t(1)   : s(1),     a(1)>
...
<t(h-1) : s(h-1),   a(h-1)>
<t(h)   : s(h),     NaN>

Robot Trajectories (Rollouts) in RoboHive Logger

While recording Robot Trajectories using RoboHive's logger, each trajectory will form a group (Rollout0,Rollout1,...), and each variable being logged will make a dataset(sensor, action,...)

Rollouts_name: {
    group{Rollout0: dataset{sensor:v1}, dataset{action:v2}, ...}
    group{Rollout1: dataset{sensor:vx}, dataset{action:vy}, ...}
    ...
}

Note that this is a special case of RoboHive logger as:

1. Dataset names (sensor, action) are the same across Rollout. RoboHive logger can have different names across different groups in its most generic case.
2. All datasets will be of similar sizes as robot trajectories are time-aligned. RoboHive logger can have different dataset lengths within and across different groups in its most generic case.

Logger Usage:

1. Record rollouts

Record grouped datasets as they are generated

from mj_envs.logger.grouped_datasets import Trace as Trace
import numpy as np

# create logger
trace = Trace("TeleOp Trajectories")

# start a new Rollout
for i_rollout in range(num_rollouts):

    # create group for the Rollout
    group_key='Rollout'+str(i_rollout);
    trace.create_group(group_key)

    # record Rollout
    for t in range(10):

        # add values to the dataset
        datum_dict = dict(
                    time=0.001*t,
                    observations=np.rand.uniform(10),
                    actions=np.rand.uniform(5),
                    rewards=np.rand.uniform(1),
                )
        trace.append_datums(group_key=group_key,dataset_key_val=datum_dict)

    # Verify length of all datasets and save logs
    trace.save("demo_rollout.h5", verify_length=True)

2. Playback Rollout

Playback the rollout again in an open loop on the environment.

3. Render logs

Render the rollout to visualize the outcome.