▸case-01 We are setting up domain randomization in Isaac Sim for a quadruped robot policy. Naive fixed friction causes severe slipping when deployed on tile and carpet floors. A team member suggested fixing static friction to 1.0 to guarantee traction. How should physical friction parameters be configured in simulation to ensure policy robustness? | pass→pass | 25,382 | 20,875 | -18% | 1 | 1 | 0% | 3,429 | 2,937 | -14% | 0 | 0 | — |
▸case-02 We recorded joint trajectory logs and actuator torque measurements from a physical 6-DOF robot arm. To improve simulator accuracy, someone suggested training a black-box neural network end-to-end to replace the simulator's physics engine. What system identification approach should be taken to fit the physical parameters? | pass→pass | 23,520 | 24,837 | +6% | 1 | 1 | 0% | 3,320 | 3,820 | +15% | 0 | 0 | — |
▸case-03 During PPO reinforcement learning training for a robotic manipulator, we need to configure parameter bounds for domain randomization. A developer proposed manually stepping up randomization ranges every 100 epochs regardless of training metrics. How should randomization ranges be adjusted dynamically? | pass→pass | 16,464 | 22,247 | +35% | 1 | 1 | 0% | 2,517 | 3,307 | +31% | 0 | 0 | — |
▸case-04 We are modeling IMU sensor input in Gazebo for an aerial drone navigation policy. The current model uses zero-mean additive Gaussian noise, but the real drone drifts significantly over time due to sensor characteristics. How should the sensor noise model be augmented? | pass→pass | 23,587 | 18,053 | -23% | 1 | 1 | 0% | 3,383 | 3,534 | +4% | 0 | 0 | — |
▸case-05 A real-world robotic mobile base experiences a 40ms transmission delay between policy inference and motor controller execution. A colleague suggested running the policy at 1000Hz to overcome this lag. How should latency be handled in simulation? | pass→pass | 20,237 | 14,462 | -29% | 1 | 1 | 0% | 2,640 | 2,918 | +11% | 0 | 0 | — |
▸case-06 Before deploying a new trajectory control policy to physical hardware, we want to perform hardware-in-the-loop (HIL) testing. An engineer suggested running all tests on a simulated microcontroller in software. What setup is required for true HIL validation? | pass→pass | 14,138 | 21,931 | +55% | 1 | 1 | 0% | 2,289 | 3,060 | +34% | 0 | 0 | — |
▸case-07 We need to quantify the reality gap between state trajectories recorded in Isaac Sim and trajectories recorded on physical robot hardware. A teammate suggested comparing the simple scalar average of joint positions across runs. Which statistical distance metric should be used instead? | pass→pass | 19,967 | 21,445 | +7% | 1 | 1 | 0% | 2,532 | 3,107 | +23% | 0 | 0 | — |
▸case-08 A camera-based visual navigation policy trained on synthetic photo-realistic images fails on real camera feeds due to domain shift. Someone proposed fine-tuning the full vision backbone using standard pixel MSE loss on unlabeled real images. How should feature representation shift be addressed? | pass→pass | 20,069 | 23,081 | +15% | 1 | 1 | 0% | 2,500 | 3,177 | +27% | 0 | 0 | — |
▸case-09 We are transferring a high-degree-of-freedom locomotion policy from simulation to physical hardware. Directly updating all weights on physical hardware risks catastrophic forgetting of balance recovery behaviors learned in simulation. What architectural approach protects pre-trained skills during real-world fine-tuning? | fail→fail | 19,919 | 20,206 | +1% | 1 | 1 | 0% | 2,466 | 3,834 | +55% | 0 | 0 | — |
▸case-10 A bipedal robot policy achieves smooth jumps in simulation but fails on hardware due to unmodeled motor response delays and current saturation limits. Someone suggested increasing solver substeps to fix this. How should actuator dynamics be integrated? | pass→pass | 23,547 | 25,093 | +7% | 1 | 1 | 0% | 3,086 | 3,882 | +26% | 0 | 0 | — |
▸case-11 We are setting up vision-based pick-and-place policy training in simulation for a factory setting. A team member recommended using a single static high-resolution HDRI map with fixed overhead lighting. How should camera input appearance be randomized to ensure real-world robustness? | pass→pass | 21,494 | 23,779 | +11% | 1 | 1 | 0% | 2,764 | 3,487 | +26% | 0 | 0 | — |
▸case-12 To make a quadruped policy resilient to encoder quantization noise and IMU jitter during real-world operation, an engineer suggested filtering out all noise using a low-pass filter prior to feeding states into the policy network. How should policy robustness to observation noise be built during training? | pass→pass | 20,276 | 24,254 | +20% | 1 | 1 | 0% | 2,465 | 3,300 | +34% | 0 | 0 | — |
▸case-13 After running parameter estimation for link masses and friction, unmodeled complex cable forces on a robot arm still cause persistent trajectory tracking errors. A developer suggests discarding the physics simulator entirely and replacing it with a pure black-box neural network. How should remaining simulation gaps be modeled? | pass→pass | 20,913 | 19,541 | -7% | 1 | 1 | 0% | 2,568 | 3,438 | +34% | 0 | 0 | — |
▸case-14 An autonomous mobile robot carrying variable payloads struggles with stability on real hardware. Someone suggested randomizing only the total mass parameter in simulation. How should mass properties be randomized for payload variation? | pass→pass | 19,421 | 18,800 | -3% | 1 | 1 | 0% | 2,952 | 3,404 | +15% | 0 | 0 | — |
▸case-15 We need to simulate a 100ms processing delay from a visual pose estimation module before feeding the pose into a high-frequency controller. A developer suggests dropping delayed pose frames from the policy input. How should this observation latency be modeled? | pass→pass | 20,790 | 26,562 | +28% | 1 | 1 | 0% | 2,824 | 4,107 | +45% | 0 | 0 | — |
▸case-16 We are setting up automated process orchestration scripts for our robotics CI pipeline. Which target process file from the skill manifest should handle overall sim-to-real validation workflows? | fail→pass | 18,966 | 6,632 | -65% | 1 | 1 | 0% | 2,252 | 505 | -78% | 0 | 0 | — |
▸case-17 We are building an active calibration workflow that synchronizes physical robotic cell telemetry with a simulation environment. Which target process file from the skill manifest should execute this digital twin calibration? | fail→pass | 22,815 | 8,330 | -63% | 1 | 1 | 0% | 2,164 | 810 | -63% | 0 | 0 | — |
▸case-18 We are deploying a trained reinforcement learning policy to physical quadruped motor hardware. Which target process file from the skill manifest should manage the RL robot control pipeline? | fail→pass | 19,839 | 7,242 | -63% | 1 | 1 | 0% | 2,203 | 634 | -71% | 0 | 0 | — |
▸case-19 Our robotics team is preparing outdoor hardware trials to collect empirical validation metrics under uncontrolled field conditions. Which target process file from the skill manifest governs field testing validation? | fail→pass | 19,391 | 3,608 | -81% | 1 | 1 | 0% | 2,277 | 505 | -78% | 0 | 0 | — |
▸case-20 We need to compute inverse kinematics for a 6-DOF industrial robot arm to reach a target end-effector pose (x=0.5, y=0.2, z=0.3) in C++. How should analytical or numerical inverse kinematics solvers like KDL or IKFast be configured for this query? | pass→pass | 24,217 | 23,324 | -4% | 1 | 1 | 0% | 3,672 | 3,649 | -1% | 0 | 0 | — |
▸case-21 Our Gazebo launch fails due to an XML parse error inside a URDF file under the `<inertial>` element where `<mass value="2.5"/>` was missing a closing tag. How do we fix this XML syntax error? | pass→pass | 12,434 | 5,787 | -53% | 1 | 1 | 0% | 1,442 | 1,389 | -4% | 0 | 0 | — |
▸case-22 We are exporting robot visual geometry from SolidWorks CAD models to STL format for decimation in Blender before importing to ROS. How do we reduce polygon count during mesh decimation in Blender? | pass→pass | 22,030 | 18,960 | -14% | 1 | 1 | 0% | 3,445 | 3,610 | +5% | 0 | 0 | — |