Research digest, week of 9 November 2026
Eight arXiv cs.RO papers on edge SLAM timing, humanoid e-stops, GPU MPC solvers, camera exposure in policy tests, time-offset calibration, retrieval baselines for VLAs, motion tracking and print path planning.
Two themes run through this week’s picks: measuring what a system actually does, and the gap between a working demo and a safe deployment.
Three papers are careful measurements you can reuse. The Raspberry Pi 5 study gives per-stage timing for RTAB-Map and ORB-SLAM3 and shows why trajectory error means little without tracking coverage. The SO-101 lighting study finds that automatic exposure hid an 11.5 times drop in room light from four imitation-learning policies. The kNN paper shows that replaying the nearest demonstration comes within a few points of current VLAs on LIBERO.
The humanoid e-stop paper is the one to read if you run legged robots near people. On a Unitree H1-2, an e-stop removes balance control and the robot falls, so stopping is not automatically the safe action.
The rest are tools. FAME-QP brings linear MPC on a humanoid within a 2 ms budget on a desktop GPU. The Galilean-group formulation folds sensor time offsets into the pose. BlinkMimic trains a G1 motion tracker in minutes, in simulation only. The additive manufacturing planner tunes its own cost weights against targets and was run on a UR10e.
RGB-D SLAM on the Edge: Benchmarking and Profiling RTAB-Map and ORB-SLAM3 on the Raspberry Pi 5 #
The authors ran RTAB-Map and ORB-SLAM3 on a Raspberry Pi 5 on six TUM RGB-D sequences, each at three resolutions, and recorded per-frame time, memory use, CPU load, temperature, trajectory error and how much of each sequence stayed tracked, and profiling each system down to function level. With RTAB-Map set to process and retain every frame, its full pipeline took 150 to 462 ms per frame against 26 to 62 ms for ORB-SLAM3, and most of RTAB-Map's time went into optimising a pose graph that grows without limit under its default memory settings. At quarter resolution both systems lost tracking on most sequences, and runs that lost tracking reported lower error on the poses they kept, so error must always be read alongside coverage. This is a dataset benchmark on the board rather than a flight test, and repeating an identical run changed trajectory error by as much as 27%, which leaves the accuracy comparison inconclusive even though timing repeated within 1% to 5%.
Runtime Safety and Emergency Stops on a Full-Size Humanoid: Deployment Lessons and a Case for Graduated Force Authority #
The paper describes the safety layer the authors built into their open-source autonomy software for a Unitree H1-2 humanoid. The layer runs apart from the control software, checks every joint command against position, velocity and torque limits, and stops the robot when the state estimate goes stale or the hard-wired e-stop is pressed or unplugged. In push tests on the physical robot with a model-predictive lower-body controller, the robot never recovered its balance, the operator hit the e-stop every time, and because stopping removes balance control the robot fell. Joint limits chosen for manipulation were also exceeded in ordinary walking and triggered the same stop. The authors argue that next to an operator or a battery pack, falling may do more harm than whatever caused the stop and propose granting force authority in stages with a named responsible party, which is a position argued from the deployment rather than a tested method.
FAME-QP: A GPU-Accelerated Quadratic-Programming Solver for Real-Time Model Predictive Control #
FAME-QP solves the dense quadratic programs that come out of linear MPC on a GPU, using an ADMM method with fused kernels replayed through CUDA graphs and an execution configuration picked by calibrating on the device. On random linear MPC problems on an RTX 3080 it had the lowest mean step time at every tested size, 3.2 to 7.3 times faster than the closest GPU QP solver, with closed-loop costs matching the other methods to within 0.06%. In a MuJoCo simulation of a Unitree G1 recovering from an 80 N torso push, it stayed under the 2 ms control period up to a 20-step horizon, while ReLU-QP went over from 15 steps. All results are in simulation on a desktop GPU, so measure latency on your own embedded target before relying on it. Code is on GitHub.
Camera equalisation of an 11.5x drop in environmental illuminance: four imitation-learning policies on an SO-101 arm #
The author trained four imitation-learning policies (pi0.5, ACT, Diffusion Policy and SmolVLA) on one shared set of 50 pick-and-place demonstrations on a physical SO-101 arm, then ran each ten times at about 520 lux and ten times at about 45 lux. Success was 20 of 40 in normal light and 19 of 40 in dim light, a difference whose 95% interval runs from minus 23 to plus 19 points, with no sign that the policies differed in sensitivity. The recordings explain why: with automatic exposure and white balance left on, mean image brightness fell by only 3% on the front camera and 19% on the wrist camera. The practical advice is to lock and report camera settings and compare image statistics with the training data, rather than quoting room lux. The sample is small, and an accidental repeat of one block matched its success count but only 6 of 10 placements.
Galilean Robots, Aristotelian Sensors: Joint Spatiotemporal Trajectory Estimation and Extrinsic Calibration #
Sensor calibration usually estimates a rigid transform and a separate time offset, and applying the offset then needs interpolation, motion compensation or a continuous-time trajectory. This paper puts the robot trajectory on the special Galilean group and the sensor extrinsics on its Aristotelian subgroup, so the time offset combines with the pose through the group operation itself. The method supports IMU-driven and twist-driven motion priors and clocks that drift, and was tested in simulation, on the Starry Night stereo camera and IMU dataset, and on a real quadrotor calibrating its IMU against motion capture and against ultra-wideband ranging. Accuracy was comparable to established methods, not better, so the benefit is a simpler problem setup rather than a lower error.
How Much of Your VLA Is kNN #
Looking up the nearest training demonstrations and replaying their action chunks, with no policy training at all, reached 91.7% success on the LIBERO simulation benchmark, where vision-language-action models now report more than 95%. Under LIBERO-Plus perturbations, retrieval degraded less than any of the policies the authors evaluated. Across nine public VLAs, they find that most of what fine-tuning does is turn the vision-language backbone into a better key for retrieval, while the actions stay close to the demonstrations. They also propose Local Action Consistency, an offline measure of whether nearby observations lead to similar future actions, which tracks retrieval success. All results are on LIBERO's simulated Franka tasks in a paper accepted to a CoRL 2026 workshop, so add a retrieval baseline to your own evaluation before crediting a policy with generalisation.
BlinkMimic: Training a Humanoid Motion Tracker in Minutes on a Single GPU #
BlinkMimic trains a single whole-body tracking policy for all 40 LAFAN1 motion clips on a simulated Unitree G1 in 3.4 minutes on one RTX 4090, averaged over three seeds and measured until a stated return threshold is met with no failed clip. It builds on earlier work that trains the policy by backpropagating through a world model fitted to parallel simulation data, adding a step size that drops once training has progressed, resets triggered by failures of the deployed policy, and GPU graph capture of each whole iteration. A leave-one-out ablation assigns each component to iteration cost, iterations needed to fit the motions, or iterations needed to clear the last failures. There is no hardware test, so treat it as a fast loop for developing tracking policies rather than evidence of transfer to a real robot. Code and per-seed results are public.
Optimization-Based Path Planning for Industrial Manipulators: Experimental Validation in Additive Manufacturing #
This offline planner for robotic additive manufacturing optimises joint paths for smoothness and self-collision avoidance alongside process costs and constraints, and a new process is added by supplying its own reference, residuals and tolerances. Long references are optimised whole or in chunks coupled at their ends, and an Iterative Weight Tuning loop sets the cost weights from the gap between each key performance indicator and its target. For carbon-fibre winding and 3D printing, the planned paths met the geometric tolerances and every KPI target and were run on a physical UR10e with joint tracking logged. The printing path ran without extrusion, so the experiment validates motion rather than part quality.