Proposed Research Plan

The first stage of the work will utilize indoor and outdoor experiments to learn and refine dynamic maneuvers for landing on vertical and sloped surfaces and on wires. The computation will be primarily off-board. Visual tracking and instrumented force plates will complement onboard sensors, helping us to characterize the limits of the latter and determine how much we can accomplish with on-board computation. We will also develop dynamic models of the plane pre- and post-contact and use these in combination with experiments to determine the envelope of initial conditions for which optimized passive mechanisms can reliably stabilize and attach the plane.

• modeling for robust control
• modeling for optimizing the mechanical systems to accommodate a wide range of initial conditions

maybe move ''learning from failure'' into year 2, make year 3 talk more about ''learning from others'' (i.e. learning from other members of a flock of planes). Q: What extra advantages accrue from having multiple planes? Maybe pick some task (e.g. environmental monitoring)?

Alexis mentions rapidly deployed array of microphones idea (e.g. to get snipers) or other array of sensors that needs to be deployed and/or reconfigured rapidly.

Motion planning idea? Formulate objective and find combination of sites that will satisfy some requirements. Network kinds of stuff...

Accumulate a quantity of stuff through multiple landings (build a nest!).

Rapidly deployed array of something in a forest... Pollution plume?

Our focus is to share information on landing... longer term application is other network ideas.

The second phase will focus on learning strategies and mechanisms to support repeated landings and take-offs in (mild) outdoor conditions. We will explore the trade space among sensors and actuators, computation and multi-functional mechanisms that facilitate take-off as well as landing. An open question is whether we can achieve repeated landings and takeoffs reliably (or at least without damage) for a fast and efficient plane with low thrust-weight ratio or whether we will be confined to planes capable of hovering briefly.

The third phase will focus on

autonomous learning from failure. Ultimately, it will sometimes prove impossible to land and attach. In this case the plane needs a safe failure strategy so that it can either abort the trial, or come around and try again, taking advantage of what it learned from the failure to refine the approach and landing. (An analogous process is seen in the %u201Ctouch and go%u201D maneuvers practiced by human pilots.) A fundamental question is the extent to which such strategies can be supported by onboard computation and how reliably the %u201Cbailout%u201D decision can be made.

Year 1

Extend the Collection of Suitable Landing Site

These things look like Developments. The first part is year 1, the later ones are maybe year 2.

Adapt the landing gear to non-vertical surfaces

Experiment with different strategy to land on non-vertical surface.

Confirm and Optimize Each Successive Funnels During the Landing Phase

Our prior work on the design and testing of the suspension of the airplane showed that landing on vertical wall can be successful for a wide range of touchdown velocities and pitch angle. This landing envelope constitutes the mouth of the first funnel to bring the airplane to a resting state on the wall and although we know the general shape of this envelope for rough surfaces [ISRR09], we are not quite sure where in the landing envelope we should plan to touchdown. To answer that question we need to obtain more information about the boundaries of the envelope (what kind of failures and how critical they are) and how the envelope shape is evolving for different surfaces. Furthermore, we need to study and optimize the aerodynamics control of the airplane (the funnel that brings the plane from normal flight at about 10 meter away from the wall to touchdown within the landing envelope), more specifically its accuracy and robustness to disturbances.

All these questions could easily be answered by using an indoor testing facility (about 10 meter long) equipped with a motion capture system and a replaceable wall surface. The motion capture system could be used to measure the exact airplane states at touchdown and easily verify the shape of the landing envelope on different surfaces. Furthermore, we could intentionally generate touchdown states that are outside the landing envelope to observe different failure mode (too high impact force, high ratio of normal-to-shear force, shear too high for the surface, one leg failure, etc) and quantify how critical the failure are and if it is possible to recover from them.

This facility would be particularly helpful in experimenting with different controllers and understanding how to optimize the landing trajectory to achieve various effects: long region of possible touchdown, high robustness to disturbance right before touchdown when flying at low speed (low controllability), landing on difficult surfaces (i.e. glass) or various maximum amount of thrust. That system would allow us to test the repeatability of the controller to land in a specific point in the landing envelop in face of uncertainty and disturbance. This measure of the landing repeatability, with the information gained about the landing envelope, would then allow us to select where in the landing envelope we would like to land to maximize the likelihood of staying away for critical failure mode in face of uncertainty or disturbance.

Take advantage of the funnels to reduce the need of sensory information and simplify control.

Also sense the response of the system to action to estimate the states.

• How to map the full state feedback controller/trajectory generator designed with reinforcement learning to a simpler controller that has similar performances?

Other ideas

• Land on glass? Land on wires?
• Shifting the center of mass to make the airplane more unstable and pitch up faster (i.e. inspired by the gecko presentation we had from Ardrian and Bob Full
• Different method of adhesion? Gripper, magnet, rat glue, etc? Paul Oh was particularly interested by landing on the side of trucks...
• Quantify the kind of disturbance we want to be able to deal with?
• Accuracy of landing?

Year 2

Enable Low Thrust-to-Weight Ratio Take-off

Create a jumping mechanism to take-off

This would allow take-off from non-vertical surfaces or take-off on a low thrust-to-weight ratio airplane. It would be interesting, if possible, to store the energy during landing for immediate use during take-off in case of unsuccessful landing.

Sensor Development for landing leading to reliable outdoor landing

From the research on funnel and what feedback is required...

Other ideas

• Funnels for take off?

Year 3

Detect and Learn from Failure

Detect and Learn from Failed Landing

We know from our experience with climbing robots that some wall are particularly difficult to cling on because of their lack of asperities. With our climbing robots, it is possible to feel these asperities with force sensors in the leg and determine if we can continue forward progress on the wall. Unfortunately, in the case of an airplane, it is difficult to detect these small asperities without first trying to land on the wall. Thus, for each landing, there is a small risk that the micro-spines won't find suitable asperities and that the plane will just slip down the wall. Although a failure to cling to the wall is not desirable, our landing system should be able to recover from it by triggering the take-off maneuver right away to allow the plane to fly away from the wall. Other types of failures could also occurs, like having a gust of wind large enough that the controller is not able to get the plane to the desired touchdown velocity or orientation. By their nature, these failures are more random and a second try might allow the plane to perch successfully.

Because of the different meanings of these failure modes, we would like to be able to extract some information while trying to land. We propose to investigate different ways to extract information during perching to categorize the landing into various levels of success and failure. To do so, propose to instrument the plane with various sensors like position and force sensors on the legs supplemented by accelerometers and gyroscopes on the body. Some of the information that can be extracted from these sensors are the force exerted on the micro-spines to get an idea of the asperities density ( could get some stuff from Alan's thesis here), the states of the plane at touchdown to know if the plane was within the landing envelope and the reaction of the plane to the wall. The reaction of the plane to the wall is particularly important as a large rebound or a large angular acceleration (i.e. one legged landing) would likely signify that the spines could release and a downward acceleration, or more generally non-steady states, would signify that the plane is sliding or tumbling down. Once these information have been gathered, a significant question remains as to how soon and how much confidence each of these measurement can provide us regarding the success/failure of the landing so that the plane can start recovering.

Although we are not quite certain at this point when would be the ideal time to recover from a failed landing, it is likely that it will be a balance between taking advantage of the natural rebound of the airplane from the wall after touchdown and the need to recover as soon as possible to prevent any tumbling of the plane down the wall. Further simulations with the simulator developed at the BDML [ISSR09] will be useful to confirm this hypothesis.

Other ideas

• Maybe also detect bad take-off. Just an idea that might need more investigation... but generally, if one leg stays stuck then it is really bad to start the propeller. Maybe we could do something like monitoring the states change (pitch angle/rate in this case) following an action, compare it to some expected states and determine if the leg is stuck or something else is going wrong and figure out an action for each failure mode.
• Determine WHEN we want to learn or explore to refine our model? When we fail? When we detect anomalies in the states-action transition?

-- MarkCutkosky - 30 Dec 2009