The Machines that Make the Machines

來源原標題

The Machines that Make the Machines

作者

Elizabeth Goodman

原始標籤

Robotics · Blackwell · NVIDIA Research · Robot Manipulation · Robotics Research and Development Digest (R²D²)

原始連結

https://developer.nvidia.com/blog/the-machines-that-make-the-machines/

摘要原文

How we taught robots to assemble GB300 tester trays and what it taught us about robot learning, mechanical intelligence, and good old-fashioned engineering The NVIDIA Grace Blackwell GB300 superchip is the engine of AI training and inference for modern foundation models. However, the process of assembling GB300 trays requires skilled physical labor in factories across the world. The NVIDIA Seattle Robotics Lab (SRL) , in collaboration with the NVIDIA Isaac engineering team, has asked the central question: Can we build intelligent robots to assemble such complex systems? The answer is yes , but the path has been challenging. Enabling robots to assemble GB300 trays is one of the hardest problems SRL has tackled in nine years as a highly-prolific research lab . The problem has also challenged our assumptions about where end-to-end learning is most powerful, and where a modular stack might be more effective. Problems at hand: Moravec returns We focused on assembling GB300 tester trays , which are used to verify the functionality and performance of GB300 compute modules prior to shipping and deployment. Based on recommendations from the NVIDIA Operations Team and Foxconn, a GB300 contract manufacturer, we zeroed in on two critical tasks: (1) busbar assembly and (2) multi-connector insertion. In busbar assembly , a long, heavy busbar (i.e., a metallic bar that transmits high current) must be grasped, transported to the tray, inserted, and screwed in at 16 locations. Fixtures and clamps must also be inserted and removed. In multi-connector insertion , two large and two small cable-mounted electrical connectors must be lifted from the tray and inserted into tight-clearance sockets. The real-world experimental setup is shown below. Figure 1. Front view of the real-world experimental setup. Robots consist of two Flexiv Rizon 4S arms and a UR10e arm with an OnRobot screwdriver Figure 2. Overhead view of the real-world experimental setup Skilled factory workers perform these two tasks reliably and fluidly, without significant mental load. However, in another example of Moravec’s paradox , the dexterity required for such tasks can be exceptionally difficult to automate. Video 1. Skilled workers at a Foxconn factory performing busbar assembly Video 2. Workers performing multi-connector insertion The first key challenge is uncertainty . Part geometry, appearance, and dynamics aren’t precisely known, and initial part poses can vary substantially. In addition, the electrical connectors are attached to cables that deform along their entire length, have significant manufacturing variation, and even change mechanically through plastic deformation and fatigue. In many manufacturing operations, uncertainty is reduced via fixed automation , where the environment is carefully controlled. Fixtures and adapters are designed to constrain parts to submillimeter accuracy, and robots can operate as repetitive motion machines. In GB300 tester-tray assembly, uncertainty must be addressed via flexible automation , as rapid design cycles and low volumes make fixed automation infeasible. Robots must be inherently adaptive and robust to uncertainty. Given the growing demand for AI infrastructure, this paradigm is a strategic priority for NVIDIA. The second key challenge is performance . In robotics research, 50%-60% success rates are often sufficient, and 80%-90% success rates can be enough to call a task solved . Cycle time, motion smoothness, and gentle handling are often secondary. In addition, statistical rigor is often insufficient; 10 to 20 trials are often executed per task, whereas reasonable confidence intervals may require far more evaluations. Our performance requirements were determined in direct collaboration with the NVIDIA Operations Team and Foxconn to ensure that solved really means solved. As of the publication date of this article, we are rapidly approaching their thresholds. Video 3. A human demonstration of what our robotic solution should perform in busbar assembly and multi-connector insertion. These steps differ slightly from those performed by Foxconn workers (shown in Videos 1 and 2); a robotic solution doesn’t require a large protective bracket and can perform screwdriving in a single phase Technical strategy: The right tool at the right time In robotics research, a premium is placed on technical novelty; this tradition promotes invention but creates little incentive to optimize baselines that may achieve higher performance. As a result, it can be easy to walk away from a research paper without understanding why prior approaches didn’t work. Our strategy for GB300 tester-tray assembly was the right tool at the right time . For each subproblem, we considered (and usually implemented) classical baselines first. When they fell short, we explored imitation learning, simulation-based reinforcement learning (RL) with sim-to-real transfer, real-world RL, and vision-language-action (VLA) models. This progression allowed us to identify the strengths and weaknesses of each approach and justify innovation and complexity. We now discuss busbar assembly and multi-connector insertion in detail; we aim for these tasks to serve as case studies for other research and engineering groups solving their own real-world manipulation problems. Assembling a busbar: Revisiting the classics Busbar assembly consists of the following steps: Grasp and insert a limit fixture (i.e., a particular fixture used to assist with assembly) into the tray Grasp a busbar (with an attached clamp) and insert it into the tray Drive 16 screws to fasten the busbar to the tray Release and remove the clamp attached to the busbar Remove the limit fixture Figure 3. The five substeps of busbar assembly The key parts in each step are boxed in red Our initial conditions were consistent with the manufacturing setting. The limit fixture and the busbar (with attached clamp) were initially placed at unconstrained locations on a flat surface next to the robot, and the screws initially rested within the busbar. The manufacturers requested us to achieve a 99.5% success rate and to complete the task in no more than twice the time of skilled workers (in this case, no more than 124 seconds). No unintended collisions were allowed with the tray, as even minor damage could be grounds for discarding the entire system. We began our work by implementing a classical pipeline (“ good old-fashioned engineering ”) consisting of separate perception, planning, and control modules, expecting a pivot to end-to-end learning afterward. As it turns out, the classical pipeline was highly effective, and no pivot was needed. The modular structure of our solution also made the system straightforward to debug and tune, as well as distribute across robot arms. For perception, we initially used NVIDIA FoundationPose , a generalist model that estimates an object’s 6D pose from its mesh, an RGB image, and a depth image. However, when an object wasn’t sufficiently visible, estimates could be off by 90 or 180 degrees. We thus extended the pipeline to ingest multiple views, produce one estimate per view, and select the estimate with the higher confidence score, which mitigated errors. (Note: We later transitioned to a specialist model called DOPER, described later.) For planning, we used a fast waypoint planner in free space to reduce our cycle times, as well as manipulation primitives based on Lissajous curves and other simple paths to align parts that may be initially misaligned. For control, we implemented a high-performance impedance controller to ensure gentle and stable contact while applying sufficient force during insertion. Whereas modern learning pipelines often implement the simplest-possible controllers, our solution leveraged automatic damping design and inertial compensation. We view high-performance control as a powerful, but underutilized tool in robot learning. To reduce cycle time, we divided work among three robot arms.