Tag Archives: box
#439875 Not So Mysterious After All: Researchers ...
The deep learning neural networks at the heart of modern artificial intelligence are often described as “black boxes” whose inner workings are inscrutable. But new research calls that idea into question, with significant implications for privacy.
Unlike traditional software whose functions are predetermined by a developer, neural networks learn how to process or analyze data by training on examples. They do this by continually adjusting the strength of the links between their many neurons.
By the end of this process, the way they make decisions is tied up in a tangled network of connections that can be impossible to follow. As a result, it’s often assumed that even if you have access to the model itself, it’s more or less impossible to work out the data that the system was trained on.
But a pair of recent papers have brought this assumption into question, according to MIT Technology Review, by showing that two very different techniques can be used to identify the data a model was trained on. This could have serious implications for AI systems trained on sensitive information like health records or financial data.
The first approach takes aim at generative adversarial networks (GANs), the AI systems behind deepfake images. These systems are increasingly being used to create synthetic faces that are supposedly completely unrelated to real people.
But researchers from the University of Caen Normandy in France showed that they could easily link generated faces from a popular model to real people whose data had been used to train the GAN. They did this by getting a second facial recognition model to compare the generated faces against training samples to spot if they shared the same identity.
The images aren’t an exact match, as the GAN has modified them, but the researchers found multiple examples where generated faces were clearly linked to images in the training data. In a paper describing the research, they point out that in many cases the generated face is simply the original face in a different pose.
While the approach is specific to face-generation GANs, the researchers point out that similar ideas could be applied to things like biometric data or medical images. Another, more general approach to reverse engineering neural nets could do that straight off the bat, though.
A group from Nvidia has shown that they can infer the data the model was trained on without even seeing any examples of the trained data. They used an approach called model inversion, which effectively runs the neural net in reverse. This technique is often used to analyze neural networks, but using it to recover the input data had only been achieved on simple networks under very specific sets of assumptions.
In a recent paper, the researchers described how they were able to scale the approach to large networks by splitting the problem up and carrying out inversions on each of the networks’ layers separately. With this approach, they were able to recreate training data images using nothing but the models themselves.
While carrying out either attack is a complex process that requires intimate access to the model in question, both highlight the fact that AIs may not be the black boxes we thought they were, and determined attackers could extract potentially sensitive information from them.
Given that it’s becoming increasingly easy to reverse engineer someone else’s model using your own AI, the requirement to have access to the neural network isn’t even that big of a barrier.
The problem isn’t restricted to image-based algorithms. Last year, researchers from a consortium of tech companies and universities showed that they could extract news headlines, JavaScript code, and personally identifiable information from the large language model GPT-2.
These issues are only going to become more pressing as AI systems push their way into sensitive areas like health, finance, and defense. There are some solutions on the horizon, such as differential privacy, where models are trained on the statistical features of aggregated data rather than individual data points, or homomorphic encryption, an emerging paradigm that makes it possible to compute directly on encrypted data.
But these approaches are still a long way from being standard practice, so for the time being, entrusting your data to the black box of AI may not be as safe as you think.
Image Credit: Connect world / Shutterstock.com Continue reading
#439119 No Human Can Match This High-Speed ...
Today at ProMat, a company called Pickle Robots is announcing Dill, a robot that can unload boxes from the back of a trailer at places like ecommerce fulfillment warehouses at very high speeds. With a peak box unloading rate of 1800 boxes per hour and a payload of up to 25 kg, Dill can substantially outperform even an expert human, and it can keep going pretty much forever as long as you have it plugged into the wall.
Pickle Robots says that Dill’s approach to the box unloading task is unique in a couple of ways. First, it can handle messy trailers filled with a jumble of boxes of different shapes, colors, sizes, and weights. And second, from the get-go it’s intended to work under human supervision, relying on people to step in and handle edge cases.
Pickle’s “Dill” robot is based around a Kuka arm with up to 30 kg of payload. It uses two Intel L515s (Lidar-based RGB-D cameras) for box detection. The system is mounted on a wheeled base, and after getting positioned at the back of a trailer by a human operator, it’ll crawl forward by itself as it picks its way into the trailer. We’re told that the rate at which the robot can shift boxes averages 1600 per hour, with a peak speed closer to 1800 boxes per hour. A single human in top form can move about 800 boxes per hour, so Dill is very, very fast. In the video, you can see the robot slow down on some packages, and Pickle CEO Andrew Meyer says that’s because “we probably have a tenuous grasp on that package. As we continue to improve the gripper, we will be able to keep the speed up on more cycles.”
While the video shows Dill operating at speed autonomously, the company says it’s designed to function under human supervision. From the press release: “To maintain these speeds, Dill needs people to supervise the operation and lend an occasional helping hand, stepping in every so often to pick up any dropped packages and handle irregular items.” Typically, Meyer says, that means one person for every five robots depending on the use case. Although if you have only one robot, it’ll still require someone to keep an eye on it. A supervisor is not occupied with the task full-time, to be clear. They can also be doing something else while the robot works—although the longer a human takes to respond to issues the robot may have, the slower its effective speed will be. Typically, the company says, a human will need to help out the robot once every five minutes when it’s doing something particularly complex. But even in situations with lots of hard-to-handle boxes resulting in relatively low efficiency, Meyer says that users can expect speeds exceeding 1000 boxes per hour.
Photo: Pickle Robots
Pickle Robots’ gripper, which includes a high contact area suction system and a retractable plate to help the robot quickly flip boxes.
From Pickle Robots’ video, it’s fairly obvious that the comparison that Pickle wants you to make is to Boston Dynamics’ Stretch robot, which has a peak box moving rate of 800 boxes per hour. Yes, Pickle’s robot is twice as fast. But it’s also a unitasker, designed to unload boxes from trucks, and that’s it. Focusing on a very specific problem is a good approach for robots, because then you can design a robot that does an excellent job of solving that problem, which is what Pickle has done. Boston Dynamics has chosen a different route with Stretch, which is to build a robot that has the potential to do many other warehouse tasks, although not nearly as optimally.
The other big difference between Boston Dynamics and Pickle is, of course, that Boston Dynamics is focusing on autonomy. Meanwhile, Pickle, Meyer says in a press release, “resisted the fool’s errand of trying to create a system that could work entirely unsupervised.” Personally, I disagree that trying to create a system that could work entirely unsupervised is a fool’s errand. Approaching practical commercial robotics (in any context) from a perspective of requiring complete unsupervised autonomy is generally not practical right now outside of highly structured environments. But many companies do have goals that include unsupervised operation while still acknowledging that occasionally their robots will need a human to step in and help. In fact, these companies are (generally) doing exactly what Pickle is doing in practice: they’re deploying robots with the goal of fully unsupervised autonomy, while keeping humans available as they work their way towards that goal. The difference, perhaps, is philosophical—some companies see unsupervised operation as the future of robotics in these specific contexts, while Pickle does not. We asked Meyer about why this is. He replied:
Some problems are hardware-related and not likely to yield an automated solution anytime soon. For example, the gripper is physically incapable of grasping some objects, like car tires, no matter what intelligence the robot has. A part might start to wear out, like a spring on the gripper, and the gripper can behave unpredictably. Things can be too heavy. A sensor might get knocked out of place, dust might get on the camera lens. Or an already damaged package falls apart when you pick it up, and dumps its contents on the ground.
Other problems can go away over time as the algorithms learn and the engineers innovate in small ways. For example, learning not to pick packages that will cause a bunch more to fall down, learning to approach boxes in the corner from the side, or—and this was a real issue in production for a couple days—learning to avoid picking directly on labels where they might peel off from suction.
Machine learning algorithms, on both the perception and action sides of the story, are critical ingredients for making any of this work. However, even with them your engineering team still has to do a lot of problem solving wherever the AI is struggling. At some point you run out of engineering resources to solve all these problems in the long tail. When we talk about problems that require AI algorithms as capable as people are, we mean ones where the target on the reliability curve (99.99999% in the case of self driving, for example) is out of reach in this way. I think the big lesson from self-driving cars is that chasing that long tail of edge cases is really, really hard. We realized that in the loading dock, you can still deliver tremendous value to the customer even if you assume you can only handle 98% of the cases.
These long-tail problems are everywhere in robotics, but again, some people believe that levels of reliability that are usable for unsupervised operation (at least in some specific contexts) are more near-term achievable than others do. In Pickle’s case, emphasizing human supervision means that they may be able to deploy faster and more reliably and at lower cost and with higher performance—we’ll just have to see how long it takes for other companies to come through with robots that are able to do the same tasks without human supervision.
Photo: Pickle Robots
Pickle robots is also working on other high speed package sorting systems.
We asked Meyer how much Dill costs, and to our surprise, he gave us a candid answer: Depending on the configuration, the system can cost anywhere from $50-100k to deploy and about that same amount per year to operate. Meyer points out that you can’t really compare the robot to a human (or humans) simply on speed, since with the robot, you don’t have to worry about injuries or improper sorting of packages or training or turnover. While Pickle is currently working on several other configurations of robots for package handling, this particular truck unloading configuration will be shipping to customers next year. Continue reading
#438613 Video Friday: Digit Takes a Hike
Video Friday is your weekly selection of awesome robotics videos, collected by your Automaton bloggers. We’ll also be posting a weekly calendar of upcoming robotics events for the next few months; here's what we have so far (send us your events!):
HRI 2021 – March 8-11, 2021 – [Online Conference]
RoboSoft 2021 – April 12-16, 2021 – [Online Conference]
ICRA 2021 – May 30-5, 2021 – Xi'an, China
Let us know if you have suggestions for next week, and enjoy today's videos.
It's winter in Oregon, so everything is damp, all the time. No problem for Digit!
Also the case for summer in Oregon.
[ Agility Robotics ]
While other organisms form collective flocks, schools, or swarms for such purposes as mating, predation, and protection, the Lumbriculus variegatus worms are unusual in their ability to braid themselves together to accomplish tasks that unconnected individuals cannot. A new study reported by researchers at the Georgia Institute of Technology describes how the worms self-organize to act as entangled “active matter,” creating surprising collective behaviors whose principles have been applied to help blobs of simple robots evolve their own locomotion.
No, this doesn't squick me out at all, why would it.
[ Georgia Tech ]
A few years ago, we wrote about Zhifeng Huang's jet-foot equipped bipedal robot, and he's been continuing to work on it to the point where it can now step over gaps that are an absolutely astonishing 147% of its leg length.
[ Paper ]
Thanks Zhifeng!
The Inception Drive is a novel, ultra-compact design for an Infinitely Variable Transmission (IVT) that uses nested-pulleys to adjust the gear ratio between input and output shafts. This video shows the first proof-of-concept prototype for a “Fully Balanced” design, where the spinning masses within the drive are completely balanced to reduce vibration, thereby allowing the drive to operate more efficiently and at higher speeds than achievable on an unbalanced design.
As shown in this video, the Inception Drive can change both the speed and direction of rotation of the output shaft while keeping the direction and speed of the input shaft constant. This ability to adjust speed and direction within such a compact package makes the Inception Drive a compelling choice for machine designers in a wide variety of fields, including robotics, automotive, and renewable-energy generation.
[ SRI ]
Robots with kinematic loops are known to have superior mechanical performance. However, due to these loops, their modeling and control is challenging, and prevents a more widespread use. In this paper, we describe a versatile Inverse Kinematics (IK) formulation for the retargeting of expressive motions onto mechanical systems with loops.
[ Disney Research ]
Watch Engineered Arts put together one of its Mesmer robots in a not at all uncanny way.
[ Engineered Arts ]
There's been a bunch of interesting research into vision-based tactile sensing recently; here's some from Van Ho at JAIST:
[ Paper ]
Thanks Van!
This is really more of an automated system than a robot, but these little levitating pucks are very very slick.
ACOPOS 6D is based on the principle of magnetic levitation: Shuttles with integrated permanent magnets float over the surface of electromagnetic motor segments. The modular motor segments are 240 x 240 millimeters in size and can be arranged freely in any shape. A variety of shuttle sizes carry payloads of 0.6 to 14 kilograms and reach speeds of up to 2 meters per second. They can move freely in two-dimensional space, rotate and tilt along three axes and offer precise control over the height of levitation. All together, that gives them six degrees of motion control freedom.
[ ACOPOS ]
Navigation and motion control of a robot to a destination are tasks that have historically been performed with the assumption that contact with the environment is harmful. This makes sense for rigid-bodied robots where obstacle collisions are fundamentally dangerous. However, because many soft robots have bodies that are low-inertia and compliant, obstacle contact is inherently safe. We find that a planner that takes into account and capitalizes on environmental contact produces paths that are more robust to uncertainty than a planner that avoids all obstacle contact.
[ CHARM Lab ]
The quadrotor experts at UZH have been really cranking it up recently.
Aerodynamic forces render accurate high-speed trajectory tracking with quadrotors extremely challenging. These complex aerodynamic effects become a significant disturbance at high speeds, introducing large positional tracking errors, and are extremely difficult to model. To fly at high speeds, feedback control must be able to account for these aerodynamic effects in real-time. This necessitates a modelling procedure that is both accurate and efficient to evaluate. Therefore, we present an approach to model aerodynamic effects using Gaussian Processes, which we incorporate into a Model Predictive Controller to achieve efficient and precise real-time feedback control, leading to up to 70% reduction in trajectory tracking error at high speeds. We verify our method by extensive comparison to a state-of-the-art linear drag model in synthetic and real-world experiments at speeds of up to 14m/s and accelerations beyond 4g.
[ Paper ]
I have not heard much from Harvest Automation over the last couple years and their website was last updated in 2016, but I guess they're selling robots in France, so that's good?
[ Harvest Automation ]
Last year, Clearpath Robotics introduced a ROS package for Spot which enables robotics developers to leverage ROS capabilities out-of-the-box. Here at OTTO Motors, we thought it would be a compelling test case to see just how easy it would be to integrate Spot into our test fleet of OTTO materials handling robots.
[ OTTO Motors ]
Video showcasing recent robotics activities at PRISMA Lab, coordinated by Prof. Bruno Siciliano, at Università di Napoli Federico II.
[ PRISMA Lab ]
Thanks Fan!
State estimation framework developed by the team CoSTAR for the DARPA Subterranean Challenge, where the team achieved 2nd and 1st places in the Tunnel and Urban circuits.
[ Paper ]
Highlights from the 2020 ROS Industrial conference.
[ ROS Industrial ]
Thanks Thilo!
Not robotics, but entertaining anyway. From the CHI 1995 Technical Video Program, “The Tablet Newspaper: a Vision for the Future.”
[ CHI 1995 ]
This week's GRASP on Robotics seminar comes from Allison Okamura at Stanford, on “Wearable Haptic Devices for Ubiquitous Communication.”
Haptic devices allow touch-based information transfer between humans and intelligent systems, enabling communication in a salient but private manner that frees other sensory channels. For such devices to become ubiquitous, their physical and computational aspects must be intuitive and unobtrusive. We explore the design of a wide array of haptic feedback mechanisms, ranging from devices that can be actively touched by the fingertips to multi-modal haptic actuation mounted on the arm. We demonstrate how these devices are effective in virtual reality, human-machine communication, and human-human communication.
[ UPenn ] Continue reading