Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Recognize Their Own Position
Transcript
- Lucas: So, we've spent a lot of episodes talking about why robot arms can't do delicate tasks — tying knots, folding shirts, opening jars. But there's a more basic problem underneath a lot of those failures, and it's something even the most expensive industrial arms struggle with: knowing exactly where they are in space at all times. Luna: You mean like proprioception in humans? The sense of where your limbs are without looking at them? Lucas: Exactly. In robotics, it's usually called 'pose estimation' — the arm's internal knowledge of its joint angles and end-effector position. Most arms use optical encoders on each joint motor to measure rotation. Those encoders give you incremental counts, so the controller can compute where the arm should be. Luna: Should be. But things drift, right? Lucas: They do. Encoders count steps, but if the arm hits an obstacle or a joint slips — say, after a collision or a power cycle — the controller loses track of the absolute position. It knows it moved X ticks, but it doesn't know the starting reference. That's why many robots have to home to a mechanical limit switch on startup. Luna: Right, that clicking noise you hear when a robot initializes — it's hitting a hard stop to recalibrate. Lucas: Precisely. But homing is slow, and it doesn't help during operation. If an arm gets bumped on the factory floor, it might be off by a few millimeters without knowing it. For a welding robot, that means a bad bead. For an assembly robot, a damaged part. Luna: So the question is: can we give robots a better internal sense of position? Something that works continuously, like human joint position sense? Lucas: That's exactly what a team at the University of Michigan addressed in a paper presented at the 2025 IEEE International Conference on Robotics and Automation. They built what they call a 'self-aware' robot arm — a six-axis collaborative arm that uses joint torque sensors and a neural network to estimate its pose without relying on encoders at all. Luna: Torque sensors instead of encoders? That's interesting. How does that work? Lucas: So torque sensors measure the force applied at each joint — both the motor torque and any external forces like gravity or contact. The insight is that if you know the arm's geometry and the mass of each link, you can infer the joint angles from the torque readings. It's like feeling the weight of your own arm to know whether it's extended or bent. Luna: But gravity depends on orientation. If the arm is horizontal, the torque at the shoulder is higher than if it's vertical. That seems like a complicated inverse problem. Lucas: It is. The Michigan team used a recurrent neural network trained on millions of torque-angle pairs. They collected data by moving the arm through random trajectories while recording both torque signals from the sensors and ground-truth angles from a high-precision motion capture system. Luna: So the network learns the mapping from torque to pose. Lucas: Yes. And they reported median position errors under two millimeters across the entire workspace — that's competitive with mid-range encoder systems. And the beauty is, it doesn't drift. Torque sensors are absolute; they don't lose their reference after a power cycle. Luna: But torque sensors are expensive, right? And they add complexity. Why not just use better encoders? Lucas: You could, but high-resolution absolute encoders — like magnetic or capacitive types — cost hundreds of dollars per joint. For a six-axis arm, that adds up. Torque sensors are also expensive, but the paper argues that for many applications, you can get dual use: torque sensing for both force control and proprioception. Luna: So it's a way to consolidate sensors. If you already have torque sensors for safe human-robot interaction, you can also use them for position estimation. Lucas: That's the idea. And it's particularly relevant for lightweight collaborative robots — the ones that work alongside people. Those arms already need torque sensing for safety: to detect collisions and limit force. Using the same data for proprioception eliminates the need for separate encoders in some designs. Luna: But there's a catch, isn't there? The neural network approach requires training on each specific arm because of manufacturing tolerances. No two arms have exactly the same torque-angle relationship. Lucas: You're right. The Michigan team trained their network on a single arm, and they tested on that same arm. When they tried it on a nominally identical arm, errors increased by about 30 percent. So there's a calibration problem — each unit might need a custom model. Luna: That could be a dealbreaker for mass production. But maybe they can build in some adaptivity? Lucas: Actually, there's follow-up work from the same group using online learning. The network continues to update its weights during normal operation, so it adapts to wear and temperature changes. That could also handle unit to unit variation if you do a short calibration routine after assembly. Luna: If today's conversation gave you something new about robotics — maybe a clearer picture of why robot arms sometimes miss their mark, or why sensor fusion matters — the way this podcast stays ad-free and independent is through listener support at buy me a coffee dot com slash fexingo. Lucas: Yeah, it's a small thing that makes a big difference. We don't run ads, and that lets us pick topics like this one — the real engineering challenges — without worrying about sponsor constraints. Luna: So if you found this useful, that's where you can help. And now back to the arm's sense of self. Lucas: One of the more interesting implications is for robots that operate in unstructured environments — like disaster response or home assistance. Those robots can't rely on homing to a limit switch because they might be dropped or flipped over. Luna: Yeah, if a robot falls over, its encoders think it's in one orientation but it's actually in another. A torque-based system would immediately detect the gravitational load change and know something is wrong. Lucas: Exactly. The Michigan team tested exactly that scenario: they powered the arm down, manually moved it to a different pose, then powered it up. The neural net estimated the new pose within seconds without any homing procedure. Luna: That's a huge deal for field robotics. It means you could have robots that power up and immediately know their configuration, even if they've been jostled during transport. Lucas: And it's not just about startup. During operation, the torque-based system can detect gradual drift due to thermal expansion or mechanical wear. Encoders don't catch that — they just count steps. But a torque sensor feels the change in the load path. Luna: So we're talking about a robot that continuously self-checks its own geometry. That's a step toward truly autonomous maintenance. Lucas: Right. And it ties into a broader trend: sensor fusion. Instead of relying on a single modality — encoders, vision, or torque — you combine them. The paper's neural network actually fuses torque data with accelerometer readings from the end-effector, which gives it a gravity reference. Luna: I've seen some work using IMUs — inertial measurement units — to supplement encoders on drones. But on robot arms, the problem is usually magnetic interference from the motors. Did they address that? Lucas: They used a MEMS accelerometer — which is solid-state, no magnets — and filtered out vibration noise with a low-pass filter. It worked well enough to improve the torque-only estimate by about 15 percent. Luna: Okay, so where does this leave the average buyer of an industrial robot? Are we likely to see torque-based proprioception in commercial products soon? Lucas: Probably not in the next two years for heavy industrial arms. Those have decades of encoder-based control invested in them. But for collaborative robots — especially new designs from smaller startups — I think we'll see it. The cost of torque sensors is dropping, and the computing power for neural nets is already onboard in many controllers. Luna: There's also the safety angle. For human-robot collaboration, you need accurate position control even if the robot is in a compliant mode. If the arm doesn't know exactly where it is, it can't guarantee a safe stopping distance. Lucas: That's a key point. Safety standards like ISO 10218-1 require that collaborative robots stop within a certain distance after a contact event. If the position estimate is off, that distance becomes uncertain. Torque-based proprioception gives a more reliable estimate because it's grounded in actual physics, not just counted steps. Luna: So you could argue that it's not just a performance improvement — it's a safety critical one. Lucas: Exactly. And that might be what pushes it into the market faster than expected. The same way that force-torque sensors became standard on collaborative arms for safety, torque-based position estimation might follow as a way to meet stricter standards. Luna: Let me play devil's advocate: if neural networks can learn the torque to pose mapping, why not just learn the inverse kinematics directly from vision? Use cameras to watch the arm and bypass joint sensing altogether? Lucas: That's definitely an active area — visual servoing. But cameras have their own problems: occlusion, lighting, calibration. If a person walks in front of the camera, or if a weld flash blinds it, you lose position. Torque sensors are immune to that. So it's really about redundancy and robustness. Luna: Alright, so the ideal system might have both: vision for external context and torque for internal state. Lucas: That's where the field is heading. The Michigan paper is one of several showing that you can combine low-cost sensors with machine learning to get high accuracy without the expense of absolute encoders. And that's important for making robots both cheaper and more reliable. Luna: It also makes you wonder: if robots can learn to feel their own position, what other senses can they learn from data that we traditionally engineered with hardware? Lucas: That's the big question. We're moving from robots that execute preprogrammed motions to robots that 'feel' their way through tasks. And it starts with knowing exactly where you are. Luna: Next episode, maybe we'll look at whether they can feel when they're about to drop something. Lucas: That's a good one. Slip detection — there's some fascinating work on that using tactile sensors. But for now, I think we've given everyone a solid look at why your robot arm might have an existential crisis after a bump. Luna: Let's just hope the next generation knows itself a little better.