Robo-MLT: Dual-System Robotic Manipulation for Long-Horizon Medical Laboratory Tasks

Zequan Zeng, Kai Wang, Delvin Ce Zhang, Huaping Liu, Zhipeng Yuan, Junqin Yang, YunLong Zhang

November 2026

TL;DR A dual-arm robot runs blood gas analysis and tube sorting by keeping the protocol as an explicit skill list, with a small multiframe monitor that advances or replans it while a VLA executes each skill. Task success is 82% and 70%, 30 and 22 points above pi0.5.

What they do

Clinical lab procedures are short but strictly ordered, and different stages can look the same: a tube at the analyzer inlet must be inserted before the assay and withdrawn after it. Robo-MLT keeps the protocol as an explicit ordered list of atomic skills with a pointer to the active one, so the stage is known from workflow progress instead of inferred from the image. It runs on a dual-arm Agilex Static-Aloha platform with three RGB cameras.

How it works

Planner. A frozen Qwen3.7-Max VLM turns the task, the protocol and the initial view into a skill sequence (7 skills for blood gas analysis, 5 for tube sorting), repeated for each tube it sees.
System 2 monitor. Frozen SigLIP2 features of the last 6 multiview frames pass through a temporal Transformer grounded in the active skill text. A completion head advances the pointer; a recovery head, or a 30 s stall timeout, triggers replanning of the remaining steps.
System 1 executor. A Qwen3-VL-2B backbone with a flow-matching head outputs 30-step action chunks for the active skill.
Asynchronous runtime. Chunks are prefetched and blended so the arm holds 20 Hz while the monitor updates at 1.33 Hz.

Results

Over 50 nominal configurations per task, success is 82% on blood gas analysis and 70% on tube sorting, against 52% and 48% for pi0.5. Sample completion is 89.00% and 92.25%.
A human advancing the pointer by hand reaches 80% and 74%, so remaining failures come mostly from execution within a skill.
A single-frame monitor drops success to 52% and 56%; a zero-shot VLM monitor gets 38% and 44%.
When an operator removes the in-hand tube, the robot recovers in 18/20 and 17/20 trials, 13/20 and 11/20 with only the timeout, and 0/20 without replanning.

Limitations

A new lab operation needs teleoperated demonstrations and retraining of the monitor. The recovery head covers only failure types seen in training, and the tested disturbance is one of them. Evaluation uses two protocols in the same workspace as data collection. Docking misalignment at the roughly 5 mm inlet and wrong-plate placements remain the main failures.

Figures

From Figure 1: Perceptual aliasing in medical laboratory automation, and the proposed Robo-MLT framework. (Top) In a ... — From Figure 1: Perceptual aliasing in medical laboratory automation, and the proposed Robo-MLT framework. (Top) In a reversible workflow, visually similar observations can require opposite actions; a policy without explicit workflow state may therefore act at the wrong stage. (Bottom) System 2 tracks the active skill from temporal observations; System 1 maps that skill and the current observation to fine-grained trajectories.

From Figure 2: Overview of Robo-MLT. A frozen Qwen3.7-Max planner maps a user command, the instrument protocol, and a... — From Figure 2: Overview of Robo-MLT. A frozen Qwen3.7-Max planner maps a user command, the instrument protocol, and an initial scene observation to an ordered sequence of atomic skills. System 2 tracks the plan with a dual-head execution monitor and invokes replanning once the active skill can no longer be completed; System 1 executes the active skill with a Qwen3-VL-conditioned flow-matching policy.

From Figure 4: The two real-world medical laboratory tasks. (Top) Blood Gas Analysis with its 7 atomic skills. (Botto... — From Figure 4: The two real-world medical laboratory tasks. (Top) Blood Gas Analysis with its 7 atomic skills. (Bottom) Blood Sample Tube Sorting: clotting and routine tubes are randomly placed on the rack and sorted into their designated plates (5 atomic skills).

From Figure 5: Real-world performance on the nominal suite. (a) Task success rate SR, (b) sample completion rate SCR,... — From Figure 5: Real-world performance on the nominal suite. (a) Task success rate SR, (b) sample completion rate SCR, and (c) completion time. Error bars denote 95% confidence intervals in (a)-(b) and 1 SD in (c).

Abstract

Reliable medical laboratory automation requires robots to complete entire workflows in protocol order while handling changes during execution. A robot may execute each motion correctly yet fail the workflow by confusing visually similar stages or stalling after human intervention. We present Robo-MLT, a dual-system framework that couples explicit workflow monitoring with learned action generation. A vision--language planner composes an ordered atomic-skill sequence from the protocol and the initial workspace observation. A multiframe execution monitor detects when the active skill has completed or become infeasible, then advances the workflow or triggers replanning. A skill-conditioned vision--language--action policy generates the manipulation trajectories, and an asynchronous runtime coordinates it with the monitor. We evaluate Robo-MLT against four policy baselines and a VLM-based monitor baseline on blood gas analysis and blood sample tube sorting, with 50 nominal configurations per method and 20 sample-removal configurations. Robo-MLT achieves task success rates of 82\% and 70\%, exceeding $π_{0.5}$ by 30 and 22 percentage points, respectively. After an operator removes the sample being handled, the system revises the remaining plan and continues processing the other samples. These results indicate that, under the evaluated conditions, explicit workflow monitoring makes long-horizon laboratory manipulation more reliable and recoverable.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong