H.O.U.N.D.
A Low-Cost Platform for aggressive off-road autonomy research
Elevation mapping for perceiving off-road environments
Rollover prevention system, tested in the real world
Integrated with high-fidelity simulator
Abstract
Off-road autonomy, crucial for applications such as search-and-rescue, agriculture, and planetary exploration, poses unique problems due to challenging terrains, as well as due to the risk involved in testing or deploying such systems. Accessible platforms have the potential to widen the field to a broader set of researchers and students. Existing efforts in making on-road autonomy more accessible have seen success, yet aggressive off-road autonomy remains underserved.
We seek to fill this gap by introducing HOUND, a 1/10th-scale, inexpensive, off-road autonomous car platform that can handle challenging outdoor terrains at high speeds. To aid development speed, we integrate HOUND with BeamNG, a state-of-the-art driving simulator to enable both software in the loop as well as hardware in the loop testing. To reduce the extent of ruggedization required, and thus cost, we integrate a rollover prevention system as a safety feature into the platform. Real-world trials over 50 kilometers demonstrate the platform's longevity and effectiveness over varied terrains and speeds.
Note that the above video is a showcase of the platform. If all of the shots in the video could have been done in autonomous mode, there would not be much left to do in off-road autonomy research
HOUND:
We provide an autonomy stack geared towards aggressive off-road autonomy. For the perception stack, we use elevation mapping with learning-based inpainting. For the high-level control, we integrate a Model Predictive Path Integral (MPPI) controller. We utilize low-cost sensors and take advantage of the Ardupilot ecosystem to handle intrinsic state estimation, hardware abstraction, and so on. The hardware is housed inside an enclosed plastic shell to protect it from environmental factors.
Simulator integration:
To aid development speed, we provide an integration with a state-of-the-art driving simulator, BeamNG, known for its crash simulation fidelity. The integration with the simulator allows both software-in-the-loop simulation and hardware-in-the-loop simulation. Software in the loop testing spoofs the perception system by providing the ground-truth elevation map and allows operating in a ROS-independent environment. Hardware in the loop simulation runs the entire autonomy stack as if it were running on the car, using ROS, with the hardware input-output being simulated by BeamNG.
Rollover Prevention System(RPS):
As rollovers are caused by inertial effects not felt outside the vehicle, they are much harder for an external operator to prevent. Unlike other works that may rely on physical ruggedization alone, we integrate a software-based rollover prevention system as a safety mechanism into the autonomy stack, to further reduce the cost of repairs.
Results:
Putting the RPS through it's paces in simulation:
We show the utility of the BeamNG simulator by using it to validate that rollover prevention system. We stress test the RPS in simulation, and find that it eliminates nearly all rollovers. On the left, we show the test with the RPS off, and on the right with the RPS on for the same scenarios.
Testing the stack in the real world:
We test the entire autonomy stack -- the physical hardware as well as the RPS, in both manual and autonomous modes, for 50 kilometers, over 4 terrains. We roll over 3 times and only damage a sacrificial part in the process.
Appendix
Design decisions
Heat management:
Sealing the body shell would cause the Jetson to reach its maximum temperature and begin throttling, creating a compute bottleneck.
We solve this by having an active negative pressure cooling system, where the fan creates a slight negative pressure in the shell, and routes the air around all the components before exhaustion. The vent is placed to reduce the chances of debris and dirt getting in.
The exhaust goes over the motor, extracting the motor's heat as well. While the exhaust air is hotter than the ambient, it is still much cooler than the motor's temperature and can provide the necessary airflow for sustained cooling.
Dealing with EMI:
The jetson, devices that use USB-3.0 protocol, and power converters, all produce electromagnetic interference that affects the GPS's accuracy. (Intel's whitepaper on USB 3.0 interference)
On big platforms, the spacing can be naturally large between such components.
To address this, we place the GPS as far away from the Jetson as possible and place one (or two) layers of Aluminum shielding between the GPS and the power electronics.
Preventing dust/dirt from getting in/on critical places:
The depth camera and the intake vent are critical points where dust and water droplets could cause harm.
While the HOUND is NOT designed to work in the rain, and should not be used on wet surfaces, occasional driving through water/wet surfaces should not immediately destroy the system
To prevent water/dust from "easily" getting into the shell, we place the intake vent and the camera in locations that maximize separation from the "water/dust splash trajectories".
For the intake vent, we also block direct paths. The vents are further shaped such that the air would have to be moving "back-to-front" in order to get into the shell.
While we have no "guarantee" that this will prevent water and dust from getting in all the time (and it probably won't), we did run it a fair bit in adverse weather conditions
The image below shows the outside after a run in early January 2023, and this video shows the insides remaining dry despite the outside of the car being "drenched" in water. Note that dust can still get in over time, but is usually a non-issue, since it does not cause short circuits and can be cleaned with pressurized air.
Note that the other holes in the main body are for passing wires to-from the chassis, in which case the wires usually block water/dust from getting through, so no special considerations are made for those gaps.