Real Robot
To use the real robots, you can submit jobs to our robot cluster. These jobs are automatically executed on a free robot. During the run, the environment transitions and performance metrics like return, success rate and reward distribution are logged and stored to files. Once the job has finished, you can download these files to evaluate the run before submitting the next job. We additionally provide a video with camera footage from three viewing angles and a visualization of the goal.
For a quick overview of the robot hardware, see About the Robot Platform.
Quick Start Guide
Create a package with your code: Structure of the User Code Package
Run the validation script: Validate Your Package Before Submitting to the Robots
Submit it to the robot: Job Submission System
Analyse the result: Accessing Recorded Data
Timing Constraints on the Real Robots
Please note that on the real robot there is a finite time budget for the execution of the policy. If this budget is exceeded, the control loop is delayed and does not run at the expected frequency. This will generally have a negative effect on the performance of the policy. In contrast to this, the simulated set-up simply waits for the next action to be ready and is therefore not sensitive to the computational demands of the policy. The time budgets were chosen as a compromise between enabling good performance of the expert policy and leaving some room for more complex models. The time budget is 10 ms for the Push task and 2 ms for the Lift task.
If the time budget is exceeded multiple times in a row, a warning is printed to user_stderr.txt (see Complete List of Generated Files). It is normal for this to happen a couple of times (around 10 times) in the beginning of a job when things are still initializing. If it happens more often, that means that the policy takes too much time to evaluate.