Atlas robotics assignment
1 like391 views
Developed the controller and mechanical research documentation on the world's most advanced humanoid - Atlas by Boston Dynamics. Different controller algorithms have been reviewed and presented in simple words. The end of the presentation contains a demo and simulation done by me.
Ad
More Related Content
What's hot (20)
Similar to Atlas robotics assignment (20)
Ad
More from Jash Shah (6)
Ad
Recently uploaded (20)
Atlas robotics assignment
- 1. Atlas HUMANOID BY BOSTON DYNAMICS The Atlas disaster-response robot made its public debut on July 11, 2013. Developed for DARPA by Boston Dynamics, where they attempted to guide the robots through a series of physical tasks representative of what might be encountered in disaster zones. Currently, it is solely used for R&D purposes. By : Jash Shah
- 2. Basics and Current Vision ● 28 Hydraulic joints with pressure sensing to mimic our veins topology. ● Lidar and Stereo vision ● 1.5m tall and 80 KGs ● Recent Focus – Implementing Parkour that involves navigating via parts other than feet as well. ● Inverse Dynamics instead of Inverse Kinematics. ● Rotate over different body axis but can not dynamically perceive the environment and implement rapid behavior creation i.e., still needs an open space.
- 3. Specific details ● They are only using the geometric decision algorithms and no real-world timing details are used !!! ● Single – Legged walk and other unique software libraries will be made public. ● It can get up after a fall in any orientation. ● The humanoid contains highly pressurized fluid that acts as its blood.
- 4. Current Applications In the DARPA competition of robotics, Atlas was able to complete all eight tasks as follows: 1. Perform Parkour skills like Backflips 2. Drive a utility vehicle at the site. 3. Travel dismounted across rubble. 4. Remove debris blocking an entryway. 5. Open a door and enter a building. 6. Climb an industrial ladder and traverse an industrial walkway. 7. Use a tool to break through a concrete panel. 8. Locate and close a valve near a leaking pipe. 9. Connect a fire hose to a standpipe and turn on a valve.
- 5. Control System One of the most interesting things in Atlas is that they use data-free models and hence, have precise model definitions and control strategies
- 6. Offline and Online Learning ● Potato Model instead of point mass with variable inertia. ● 2-step optimization process that gives shape to the entire robot. ● Servo and Regulatory Problems. ● Implementing Model Predictive Control -> Transferring the online learning to reality. ● Use Perception based techniques as well to move around ● Do not use data driven techniques for developing the physical controller to make the fairly complex model simple. But, plan to use RL in the next model. The Offline is developed from scratch and is aimed at the longer horizon. It develops a set of libraries for the bot. The Online learning is task-generic and is an extension of the offline learning, rejecting the disturbances from ideal offline learning
- 7. Model Predictive Control Model Predictive Control is a model based control that is specific to the model and decides the control parameters based on the dynamic model equations. It predicts the future state of the plant and optimizes the controller action. Also very useful in MIMO. 2 models of MPC : 1st is to solve many different tiny problems and solve them and 2nd is to keep the problem sparse and use complex algebra to solve it.
- 8. Bipedal Robot Balancing and basic Control ● For balancing, Atlas focuses on finding the Angular Excursion and ZMP of the bot. ● The zero moment point is a very important concept in the motion planning for biped robots. Since they have only two points of contact with the floor and they are supposed to walk, “run” or “jump” (in the motion context), their motion has to be planned concerning the dynamical stability of their whole body. Ankle strategy, Hip strategy and Step strategy
- 9. Advanced Locomotion on non-linear terrain Perception of edges ● Elevation maps of the environment built from 3d sensors, such as Lidars, can be used to find possible stepping regions or to obtain a guess about the upcoming foothold. ● Based on the angle measured, the flat-foot model is applied and approximate pose is determined Controller Framework ● Instantaneous Capture Point ( ICP ) and Centre of Pressure ( CoP ) are determined and based on that the bipedal motion is approximated to 2 Inverted Pendulums. ● Quadratic Programming is used to optimize the trajectory and kinematic model.
- 10. Traversing Doors
- 11. ...Contd ( door traversing) ● The entire process is thought of as an FSM and each Markov model is solved indefinitely. ● Perception of the door and traveling involves the Expectation-Maximization algorithm ● End-effector trajectories were developed using Rapid Random Trees. The approach could be broken down into following steps: ○ Approaching Handle ○ Turning Handle: ○ Pulling Door Open ○ Blocking Door from Closing: ● The environment geometry is generated by convex decomposition of point clouds.
- 12. Obstacle avoidance and Jump Control ● Many humanoid applications can be decomposed into a two stage control problem: a behavior level controller that outputs high level commands and a low level controller that is responsible for generating joint commands. In order to fully utilize the workspace and be robust to external perturbations, the low level controller has to take full body kinematics and dynamics into consideration. ● As shown in the picture, the MPC predicts the position well in advance, and the normal forces that adds to the disturbances are measured. When the actual process takes place ( right pic ), it is the online optimization that is to regulate the effects from predicted and real observables !!!
- 13. Back - Flip Control System The 2 step process is very interesting, what they do is in the first step, they take the potato model and model the whole body - motion for the momentum and in 2nd step they overlap this model to measure the joint constraints and the spatial distribution of the robot. In the end, after they have modeled the servo problem, they fine tune it for external disturbances incorporating the regulatory motion. This also includes typical cases like flipping over the edge of the box, etc.
- 14. Software Simulations - Controller and Perception Simple Walk Generator Perception and Depth Analysis
- 15. Inverse Dynamics over Kinematics and Angular Excursion ● One popular approach to controlling humanoid robots is through inverse kinematics through stiff joint position tracking. On the ● other hand, inverse dynamics that focus on the torques generated have gained increasing acceptance by providing compliant ● motion and robustness to external perturbations. This is the main research focus of Atlas. However, the performance of such ● methods is heavily dependent on high quality dynamic models, which are often very difficult to produce for a physical robot. ● Inverse Kinematics approaches only require kinematic models, which are much easier to generate in practice.