search

bulletphysics / bullet3

Bullet Physics SDK: real-time collision detection and multi-physics simulation for VR, games, visual effects, robotics, machine learning etc.
http://bulletphysics.org
Other
12.73k stars 2.88k forks source link

Setting time step to 0.001 or 0.01 makes the robot bounce off the ground and float. #1460

Closed brendantidd closed 4 years ago

brendantidd commented 7 years ago

I am trying to use pybullet for a reinforcement learning algorithm (DDPG) using a custom urdf model I have made of a biped (12 joints and torso). I initialize my environment with gravity, then p.setTimeStep(DT), and step through the simulation at each time step by sending actions with p.setJointMotorControlArray(self.Id,order,p.POSITION_CONTROL, actions), then p.stepSimulation(). I am wondering if there is a limit for DT to give it time to run the simulation? I have tried DT of 0.001 and 0.01 and the robot seems to bounce off the ground plane and then float through the air like it is in water. When I run DT of 0.1, the feet ground (though seem to sink into the ground plane) and the simulation behaves as I might expect with the robot falling quickly over. Can I run the simulation quicker? Am I missing any steps in my implementation? Thanks for your time!

erwincoumans commented 7 years ago

Never seen anything like this, it could be self-collision.

Am I missing any steps in my implementation?

I'm not a fortune teller or clairvoyant. If you can provide a simple reproduction case (a pybullet file and URDF etc), someone may look into it.

Usual delta time values for robotics ranges from 0.01 sec (pretty big) to 0.001 (1khz). I wouldn't recommend using a bigger timestep, but smaller timestep are better (but slower). The default timestep is 240 Hz.

brendantidd commented 7 years ago

ah yep, sorry about that. I appreciate you taking the the time to respond.

This repository contains my urdf under 'assets', I created it from a xacro using a ROS command. my gym-like environment is called 'env.py'. The rest of the code is from OpenAI baselines: https://github.com/neveryoubetray/ddpg

Thanks again.

ojasjoshi commented 4 years ago

@erwincoumans I see that you are mentioning delta time in Hz. Does that mean that the real-time-factor(RTF) for pybullet simulations is always(or by default) = 1. Or is the update rate(and hence the RTF) configurable via maxSolverIterations parameter?

erwincoumans commented 4 years ago

This topic is unrelated to 'real-time factor'. Closing out-of-date issue.