Latest / Embodied AI 101 / YOR: Your Own Mobile Manipulator for Generalizable Robotics

Transcript

YOR: Your Own Mobile Manipulator for Generalizable Robotics · Embodied AI 101

• YOR: Your Own Mobile Manipulator for Generalizable Robotics. Robotics research has lately rushed toward data-driven intelligence – learning from tons of real interactions to approach human-level performance. But hardware remains a bottleneck. Most capable mobile manipulators (think Fraunhofer’s TIAGo, Yaskawa’s Motoman, or Persistent Systems’ Fetch) cost tens or hundreds of thousands of dollars, putting them out of reach for even well-funded labs. At the same time, actuator commoditization (cheap motors, controllers, sensors) has opened the door to low-cost robots. A key question is: how do you get the maximum dexterity, reach, and flexibility when you only have a few thousand dollars to spend?. YOR (short for Your Own Robot) tackles this head-on. Introduced in a recent 2026 arXiv paper, YOR is a fully open-source mobile manipulator built around a ~$10K parts budget. Its creators (the team led by Anjaria et al. at NYU and collaborators) cobbled together an omnidirectional wheeled base, a vertical lift tower, and two lightweight 6-degree-of-freedom arms. The result is a platform that achieves whole-body mobility, bimanual grasping, and vertical reach for overhead tasks – all at a price point one or two orders of magnitude below commercial alternatives. The authors emphasize three core contributions: (1) a practical form factor that balances affordability and dexterity without the complexity of legs or humanoid limbs, (2) a fully open-source hardware and software release (complete CAD models, bills-of-materials, control code, etc.), and (3) empirical demonstrations of whole-body teleoperation, learned manipulation policies, and autonomous navigation in cluttered indoor scenes. In a nutshell, YOR makes advanced mobile manipulation accessible to any lab or hobbyist who wants to “build their own robot.”. Below we dive into how YOR is built, how it’s controlled, and what it can actually do. We’ll see how the team leveraged off-the-shelf components and clever design to hit that sub-$10K target, how they collect data via human teleoperation and train policies, and how YOR handles full SLAM-and-navigation autonomy. Along the way we’ll touch on related efforts (like Stanford’s TidyBot++), and discuss why an open, affordable platform is exciting for robotics research. Throughout, we’ll cite directly from the paper and project materials to give specifics on performance and design choices. Design and Hardware Architecture. At its core, YOR is a modular mobile manipulator, breaking the robot into three main subsystems: an omnidirectional wheeled base, a telescoping vertical lift, and a dual-arm manipulator “torso.” This combination is deliberate. Holonomic (full-motion) base wheels let the robot translate and rotate independently, crucial for maneuvering in tight indoor spaces without constant reorientation. The telescoping lift places the shoulders on a column that can raise or lower almost 63 cm, meaning the arms can reach the floor or high shelves without contorting their joints. And two arms (instead of one) give true bimanual dexterity: handing objects from one gripper to the other, or stabilizing a large item with one while manipulating with the other. Throughout, every part is chosen for low cost and easy sourcing. As the authors explain, YOR’s design is guided by four principles: low cost and modularity, precise controllability, safety/compliance for human environments, and open-source accessibility. For instance, unlike highly custom robots, YOR is built almost entirely from commercial off-the-shelf (COTS) parts. The wheeled base uses four REV Robotics “MAXSwerve” modules (3-inch omni-­directional wheels, each driven by a pair of brushless motors for steering and driving). This swerve drive makes the base fully holonomic – it can move sideways, diagonally, spin in place, or any combination, without needing to turn first. By contrast, a typical differential-drive base would have to perform back-and-forth maneuvers in tight spots. The dense layout of the swerve modules, along with the base’s aluminum-framed chassis, keeps the footprint compact while providing stability. In fact, the team deliberately keeps heavy parts (battery, compute unit) low in the chassis to put the center of mass near the floor, improving balance when the arms swing. Above the base sits a vertical lift column adapted from an off-the-shelf standing-desk actuator. This telescoping assembly raises the robot’s “shoulders” from about 0.6 m when fully down to 1.24 m at full extension. Since even modest manipulators have limited reach, this column effectively doubles the workspace: YOR can reach objects on the floor or on high counters without the arms contorting. The lift itself is custom-controlled by a small onboard microcontroller (a Raspberry Pi Pico) which reads a quadrature encoder and drives a DC motor via an H-bridge, giving precise position and speed control (up to ~35 mm/s lift speed). Importantly, the lift mechanism is self-locking (a typical desk actuator feature), so it holds its position if power is cut – a safety feature to prevent unexpected drops when YOR is idle or if power fails. Flanking the lift are two identical 6-DoF robotic arms made by AgileX (the PiPER arm). Each PiPER arm costs about $2,500 and provides human-like dexterity (shoulder, elbow, wrist joints). The arms are mounted on a custom shoulder plate angled slightly downward, which serves two purposes: it gives the arms better reach toward the floor and prevents their “elbows” from colliding by angling them outwards. Each gripper is custom too – the team built simple angular-jaw grippers with large opening diameter, and they cleverly incorporated smartphones (iPhones) as built-in vision sensors. In other words, YOR’s manipulators have compact RGB cameras where the fingertips are, providing close-up views of grasped objects. This use of consumer hardware (phones for vision, gaming controllers for input, etc.) is emblematic of the project’s frugal mindset. Overall, the lift plus two arms arrangement provides true whole-body mobility: the base can reposition, the lift can adjust height, and the arms can cooperate in 3D. On the software/control side, YOR enforces modularity: each subsystem (base, lift, each arm, sensor suite, compute) can be upgraded independently. For example, the arms use a compliant control scheme. A simple joint-space stiffness controller (much like a gravity-compensated PD controller) runs at 200 Hz on the arms’ controllers. This lets the arms be “soft” to external forces, which is safer around people, while still tracking desired joint trajectories. The robot also supports full operational-space control: one can move an end-effector in XYZ space and the controller will compute inverse kinematics (using Minkowski-based algorithms) to achieve that pose. Critically, YOR’s whole-body control framework lets the operator or planner command any cross-section: you can drive the base, move the lift, and command each arm either by joint angles or by target poses. For teleoperation, the team integrated a consumer VR tracking system (Meta Quest controllers and headset). The humans wear a VR headset and hold two controllers; real-time positional tracking maps the left and right controller poses to the robot’s left and right end-effector target poses, respectively. At the same time, pushing a button on the controller raises or lowers the lift. In this way, an operator can naturally “teleoperate” the entire robot: walking around, grabbing vision images, and moving the robot as if it were an avatar. This design makes demonstrations intuitive and much faster to collect than, say, keyboard-and-joystick interfaces. Taken together, YOR’s hardware strikes a balance that no other sub-$10K robot hits: it has holonomic mobility, vertical reach, and bimanual dexterity in one package. The authors even include a figure comparing YOR’s workspace and price to other low-cost platforms, highlighting that YOR’s combination of features outclasses common alternatives in that price range. For example, many cheap platforms either skimp on mobility (using skid-steer wheels) or omit any lift, or have only a single arm. YOR aims to pack all those capabilities in by cleverly reusing proven parts. Whole-Body Teleoperation and Data Collection. Because YOR is designed to leverage data-driven learning, the team emphasized easy data collection. After assembling the robot, they built a teleoperation setup that lets a human quickly demonstrate tasks in a real home or lab setting. Recall the Meta Quest controllers: the system maps the tracked controller poses to the robot’s wrist poses in its base coordinate frame. There’s also a user interface for base driving: the touchpads or joysticks on the headset can generate planar velocity commands for the robot base. The vertical lift is controlled by holding a grip button (one axis of controller for up/down). In effect, the human pilot can naturally command arm motion (with 7D poses + gripper open/close) for each arm and base motion (vx, vy, ω) all at once. To prevent operator fatigue, the authors decoupled height control (lift) from arm control so you can move your hands level while pushing on triggers for lift up/down. This teleoperation framework proved effective. The paper shows snapshots of YOR being teleoperated to do household chores: loading a dishwasher rack with dishes, watering potted plants, picking objects off the floor, etc. We see images of YOR bent over a sink, one hand filling a cup, another placing a plate. These scenes illustrate that even though YOR is low-cost, it can handle fairly complex everyday tasks when guided by a human. This human-in-the-loop demo collection is critical, because it generates the high-quality data needed for learning. The authors intentionally built YOR and the teleop system for out-of-lab, “in-the-wild” use – no specialized motion-capture or safety fences. Collectively, YOR’s design and teleop stack create a pipeline where a single operator can amass many hours of demonstration data in realistic environments. The team then leveraged this pipeline to collect data for imitation learning. In one key experiment, they seated YOR in a realistic kitchen-like setting with furniture and obstacles and defined a pick–carry–place task: grab a bottle from a table, carry it through the room, and drop it into a bin. This task required YOR to coordinate all parts of its body: both arms had to cooperate (one to stabilize the object, one to lift it), the lift column had to extend to pick up from the table, and the base had to navigate around an obstacle. Using the teleop interface, an experienced operator provided 100 successful demonstrations at 30 Hz control frequency. Each demo included RGB camera images from both wrist cameras and the upward-facing ZED 2i “head” camera, as well as YOR’s proprioceptive state (end-effector poses, lift height, base odometry). These demonstrations were fed into a sequence model: specifically, a VQ-Behavior Transformer (VQ-BeT). In plain terms, this is a behavior-cloning policy that uses a transformer neural network to map from recent visual + state context to the robot’s next action. The visual inputs (first-person views from each wrist and the head camera) were passed through a shared ConvNet encoder (ResNet-50 pre-trained in a self-supervised way), and the robot state was linearly projected; their features were concatenated into a sequence. The transformer decoder then outputs a discretized action vector at each timestep: this action includes where to move each wrist next (7D pose per arm), where to steer the base (vx, vy, cos(yaw), sin(yaw)), how high to move the lift, and gripper open/close commands (about 19 dimensions total). In short, it learns to reproduce the teleoperated behavior in a single shot. When they ran the learned policy on YOR in the same environment, the results were impressive. The policy achieved 10/10 success in picking up the object and lifting it to the correct height, and 9/10 overall success over the full pick-carry-place sequence. (The single failure was due to camera occlusion and resulting odometry drift, not a grasping error.) In other words, after around 100 demos, YOR could autonomously pick the bottle, avoid obstacles, navigate, and drop it in the bin almost every time. This is a testament to both the data quality and YOR’s repeatability. The paper illustrates one representative successful trajectory, showing the robot’s arms grasping the object, turning around, lifting the lift, and finally dumping into the bin. These learning experiments highlight a key point: YOR’s form factor makes such complex multi-modal policies feasible on limited data. Coordinating base + lift + two arms in a cluttered space is a tough task, but by giving the human operator natural 3D control and capturing that data, the policy could generalize to new runs. In many older systems, collecting 100 high-dimensional demos might be prohibitive due to hardware limitations; here it’s straightforward. The authors suggest that this workflow – teleoperate many tasks, train large transformer policies – could be repeated for all sorts of “everyday” activities (e.g. pick-and-place, tidying, kitchen chores) in real homes. Because the hardware is modular and cheap, multiple YOR units could even be run in parallel for larger datasets, unlike commercial platforms. The learned policy uses all aspects of the robot’s body, showing that YOR truly achieves whole-body coordination. Autonomous Navigation and SLAM. Beyond teleop and imitation learning, YOR was also equipped with a complete onboard navigation system to test autonomous movement in cluttered environments. For sensing, the robot carries a ZED 2i stereo camera on its head, which provides synchronized RGB and depth, plus an IMU for fast motion estimation. Onboard software runs a state-of-the-art visual-inertial SLAM (based on VINS-Mono algorithms) to continuously estimate YOR’s pose in space. This VO (visual odometry) data, fused with the stereo depth measurements, feeds into a dense voxel map of the world. In practice this means that as YOR moves, it stitches together a 3D map of the floor, obstacles, and walls in its environment. The system even does simple semantic processing: for example it detects the floor plane and inflates it in the map to avoid “cliffs.”. Navigation planning is fairly classic: the voxel map is projected into a 2D cost grid, then A* search finds collision-free paths to a goal. A pure-pursuit controller with PID follows this path using the swerve-drive constraints. The twist here is the “whole-body” aspect: path planning is only one part – the arms might need to stay at a fixed pose while driving. The YOR control stack supports that by having a separate loop: if you tell the robot to hold its current end-effector pose while the base moves, the low-level controller computes compensatory arm motions so the gripper stays world-frame stable. In effect, YOR can carry an object in its gripper while driving without letting it slip, by doing real-time kinematic compensation. To validate this, the authors set up a few interesting tests. In one “tally-mark” demo, YOR held a pen in its gripper at a fixed position above paper and drove around in a loop (like tracing a circle on paper while moving). After 10 loops, the marks it made formed a tight cluster with a radius under 12 mm. This shows that the odometry and control are quite precise: the pen returned very close to its start point each time. In another test, YOR picked up a ball, held it steady, and translated the base; the arm controller successfully canceled out base motion so the ball stayed fixed in world coordinates, with errors under 16 mm. These are pretty tight tolerances for a system on a $10K budget. Finally, they tried dynamic obstacle avoidance: a person walked in YOR’s path and YOR had to replan on the fly. The system updated its voxel map and A* plan within about 1 second, smoothly detouring around the moving person. A replicated plot (Figure 10 in the paper) shows YOR smoothly curving away from the human and then back on track, demonstrating robust SLAM-based navigation. In summary, YOR handles full-loop autonomy in indoor scenes. It localizes with visual SLAM, builds maps, plans paths with A*, and even concurrently manipulates or holds objects while moving. The success metrics (12 mm repeatability, 1 s replan on dynamic obstacles) show that even though YOR costs far less than research-grade rigs, it delivers solid navigation performance. This positions YOR not just as a “toy”, but as a capable mobile manipulator: it can be set as an experiment to run without a tethered supervisor, e.g. to autonomously perform chores or fetch tasks. Imitation Policy Learning on Bimanual Tasks. We already mentioned the learning experiment for the recycling task via VQ-BeT. To expand on that, let’s look at how YOR supports generalizable policy learning. The goal here is to train controllers that can handle varied real-world scenarios, beyond that one fixed task. A few elements contribute: multiple camera views, tactile feedback (to some degree, via stiffness sensing), and multi-arm coordination. The policy’s input (“observation”) fused images from the left wrist, right wrist, and head cameras, plus proprioception (joint angles, base odometry, lift height). This rich multi-modal input is crucial for generalization: e.g. if one camera view is occluded, others or the proprio history can help. The transformer model itself (VQ-Bert) was pre-trained on large image datasets (for the ResNet encoder) and then fine-tuned with the collected demos. By discretizing actions in a codebook, it avoids having to regress continuous velocities, simplifying learning. While the published experiments reported above used ~100 demonstrations, YOR’s open-source nature means you could scale this up hugely. Imagine many labs worldwide each recording important tasks (kitchen clean-up, laundry folding, desk tidying) with YOR – the eventual pool of data could train a giant, versatile policy or even a foundation model for mobile manipulation. The authors explicitly suggest YOR paves the way to “large-scale, real-world data collection” for mobile robots. Right now, learning research has often been done on stationary arms (no navigation) or tiny robots; YOR opens the door to corpuses of demonstrations that include whole-body locomotion and bimanual action. One subtle point: the learning studies here were done offline (behavior cloning from demos). The authors note that YOR could also enable on-policy or offline reinforcement learning in the future. Because it’s open and cheap, one could even run multiple YORs in parallel on the same task distribution to collect more varied trajectories. In terms of “generalization”, it’s telling that the paper frames YOR as an enabler of zero-shot transfer and perceptive planning research. With suitable training, one hopes to move beyond single tasks and get robots that can adapt to new objects or rooms, using vision. YOR is positioned as a platform for exactly that kind of embodied AI research. A related recent project is TidyBot++ – an open-source holonomic mobile manipulator from Stanford/Princeton. TidyBot++ also uses a fully driveable base and a dual-arm setup, collecting household task demos via teleoperation. Its authors describe it as “an open-source design for an inexpensive, robust, and flexible mobile manipulator” with a base using powered casters for true omnidirectional motion. YOR shares this spirit but adds a vertical lift and a focus on aggressive cost-down. In fact, YOR’s authors compared their design to TidyBot++ and other small robots (like XLeRobot and a robot called RB-Y1) in terms of workspace and dexterity. Their claim is that no other <$10K platform matching YOR’s combination of features even exists. Thus YOR doesn’t just repeat past designs; it pushes the envelope by integrating more subsystems (lift + bimanual) within the tight budget. Impact, Accessibility, and Future Directions. What does YOR really promise for the robotics community? The headline is lower barrier to entry. By publishing all CAD files, detailed build instructions, and software under open licenses, the authors make it genuinely possible for any researcher (with some engineering hustle) to assemble their own YOR clone. For smaller labs or universities without deep pockets, this could revolutionize what experiments are feasible. Instead of one expensive robot, a lab might build three YORs and run parallel experiments (say collaborative sorting tasks) that were never possible before. Furthermore, because the design is modular, groups can customize YOR for their needs: swap in a better sensor, upgrade the arms, or even attach custom tools, without redoing everything. This flexibility means YOR could spawn a whole ecosystem, much like how drones share modular avionics. The authors stress that diverse users will “build, modify, and improve” YOR in different ways. Over time, the community might innovate on YOR – perhaps adding 7th joints on the arms, or better grippers – and share those back as improvements. The practical upshot is broad generalization. YOR removes the „deployment barrier,“ enabling extensive real-world testing of mobile manipulation algorithms. For example, a person might train a policy in their home and evaluate it in a different building, testing zero-shot transfer. Or multiple YORs could roam an office to gather a variety of encounter data for training safety protocols. The paper even envisions YOR powering research into “perceptive loco-manipulation” – combining vision, planning, and whole-body control on-the-fly. All these research directions have been bottlenecked by hardware access: YOR’s existence lowers that bottleneck. One concrete takeaway the authors highlight is that affordable platforms can indeed deliver precision and dexterity. They report that YOR’s sub-centimeter consistency in repetitive tasks rivals what you’d expect from pricier systems. The lift and arms, despite being inexpensive, still reach typical household heights and payloads. In their conclusion, they succinctly note that YOR “demonstrates that mobile manipulation platforms can simultaneously deliver dexterity, vertical reach, bimanual manipulation, omnidirectional motion, and compliance at a low cost”. That’s a new baseline: where once such a robot might have been $50K, we now have a $10K solution that “establishes a new baseline for affordable bimanual mobile manipulation,” as the paper puts it. Looking ahead, the authors list obvious enhancements: 7-DoF arms (to handle awkward postures), more advanced servos, semantic mapping capabilities, etc. But even in its current form, YOR invites immediate projects. It could serve as a testbed for language-conditioned manipulation or multi-agent coordination. Educators could use it to teach robotics courses with a real platform. And importantly, because everything is open-source, results achieved on YOR should be reproducible by others – a perennial challenge in robotics research. In short, YOR’s message is that “your own robot” doesn’t have to be a pipe dream. By carefully choosing hardware and open design, the team shows that a broad set of real-world, safety-conscious tasks become tractable at a tiny fraction of the usual cost. This blurs the line between lab prototypes and practical robots. As robotics continues to lean on machine learning, platforms like YOR could be the workhorses driving the next wave of embodied AI – gathering data, testing policies, and making the vision of generalizable home robots a bit more real. Sources: The descriptions above are drawn from the YOR project’s official documentation and a recent arXiv paper (Anjaria et al., 2026) detailing the design and experiments, along with insights from a community blog summary and related literature such as the TidyBot++ release. These sources provide detailed breakdowns of the hardware, software, and test results that underpin YOR’s capabilities.