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

Anonymous Authors
Paper under review

Abstract

Reliable medical laboratory automation requires accurate manipulation and consistent execution of the operating protocol. Visually similar stages can require different actions, while human intervention can make the active step impossible to complete. We present Robo-MLT, a dual-system framework that couples explicit workflow monitoring with learned action generation for long-horizon laboratory manipulation.

A vision–language planner turns the protocol and initial workspace observation into an ordered sequence of atomic skills. A multiframe execution monitor tracks the active skill, verifies completion, and triggers replanning when the skill can no longer be completed. A skill-conditioned vision–language–action policy generates manipulation trajectories, while an asynchronous runtime coordinates workflow decisions with continuous action streaming.

We evaluate Robo-MLT against four policy baselines and a VLM-based monitor on blood gas analysis and blood sample tube sorting. Across 50 nominal configurations per task and method, Robo-MLT achieves task success rates of 82% and 70%, exceeding π0.5 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.

Method Overview

Robo-MLT separates workflow decisions from the motions needed to carry them out. A frozen Qwen3.7-Max planner reads the user command, instrument protocol, and initial workspace observation to compose an ordered plan over a predefined library of atomic skills. It orders and repeats supported operations according to the observed tube count, positions, and categories.

System 2 monitors workflow progress. A lightweight monitor uses six-frame windows of multiview observations and the active skill instruction to make two decisions. Its completion head confirms that the skill has finished and advances the plan pointer; its recovery head detects states in which the skill can no longer be completed. A stall timeout provides a backup recovery trigger. Replanning revises the remaining steps using the protocol, completed steps, attempted-skill history, and current scene.

System 1 executes the active skill. A Qwen3-VL-conditioned flow-matching policy maps the skill instruction and current observations to continuous action chunks. The asynchronous runtime overlaps chunk generation with execution, allowing the monitor to update at 1.33 Hz while the controller streams buffered actions at 20 Hz.

Robo-MLT architecture: a vision-language planner creates an ordered skill sequence; System 2 monitors completion and recovery; System 1 generates actions for the active skill.

Overview of Robo-MLT. The planner specifies the skill sequence, System 2 determines when to advance or replan, and System 1 generates the actions for the active skill. An asynchronous runtime coordinates temporal monitoring with continuous manipulation.

Contributions

  1. Explicit workflow state for long-horizon manipulation. Robo-MLT combines protocol-grounded skill planning and low-frequency workflow monitoring with high-frequency action streaming, helping distinguish visually similar stages that require different actions.
  2. Temporal monitoring for completion and recovery. A lightweight multiframe monitor verifies skill completion and detects when the active skill can no longer be completed. It advances the plan or triggers replanning of the remaining steps from the protocol and current scene.
  3. Real-world evaluation of workflow reliability. Experiments on blood gas analysis and blood sample tube sorting include baseline comparisons, ablations, offline analysis of skill-transition timing, and sample-removal tests, measuring both end-to-end task success and recovery after intervention.

Experiments

We evaluate two laboratory workflows on an Agilex dual-arm Static-Aloha platform with one global and two wrist-mounted RGB cameras. Both systems run on a single RTX 4090 GPU. All methods share the same observations, robot state, action space, control frequency, and safety limits.

Quantitative Results

The nominal evaluation compares Robo-MLT with ACT, SmolVLA, π0, π0.5, and a VLM-based monitor, using the same 50 configurations per task for every method. Task success requires all samples to be processed in protocol order within the task-specific time limit. The table highlights Robo-MLT and two key baselines.

Method Blood Gas Analysis Tube Sorting
Robo-MLT 82% (41/50) 70% (35/50)
π0.5 52% (26/50) 48% (24/50)
VLM-Monitor 38% (19/50) 44% (22/50)

Robo-MLT improves task success over π0.5 by 30 and 22 percentage points and achieves sample completion rates of 89.00% and 92.25% on blood gas analysis and tube sorting, respectively. Sample completion measures the fraction of samples that finish all required operations, including those in partially completed batches.

In the sample-removal evaluation, an operator removes the in-hand sample, and success requires autonomous completion of the workflow for all remaining available samples. Robo-MLT recovers in 18/20 trials (90%) on blood gas analysis and 17/20 (85%) on tube sorting, compared with 65% and 55% when replanning is triggered only by a timeout. Disabling workflow-level recovery yields no successful recoveries on either task.

Real-World Rollouts

Task 1 · Blood Gas Analysis

Each sample passes through seven skills: grasp the tube, remove the cap, discard the cap, dock with the analyzer, insert the tube onto the sampling probe, withdraw it after the assay, and discard it. The approximately 5 mm inlet clearance requires precise alignment. System 2 uses the plan pointer and temporal evidence to distinguish the visually similar insertion and withdrawal stages.

Task 2 · Blood Sample Tube Sorting

Clotting and routine tubes start at randomized rack locations and must be sorted into their designated plates. The planner identifies tube positions and categories from the initial observation and specifies a pick-and-place order. System 2 tracks progress across repeated, visually similar cycles.

System 2 Monitoring in Action

The clips pair synchronized camera views with the probabilities predicted by System 2. The plan advances after the completion probability exceeds τdone for two consecutive updates. Recovery is triggered after the recovery probability exceeds τrec for three consecutive updates, or when the active skill times out. The replanner then revises the remaining steps using the protocol and current scene.

Task 1 · Blood Gas Analysis
Task 2 · Blood Sample Tube Sorting

Perceptual Aliasing in End-to-End Policies

Representative end-to-end policy failure on blood gas analysis. Insertion and withdrawal can produce nearly identical views at the analyzer inlet. In this example, the policy repeats insertion when the protocol requires withdrawal, stalling the workflow. The failure illustrates why action selection depends on workflow progress as well as the current image.

Scope and Limitations

Validation covers two protocols in a single workspace. The planner can reorder and repeat existing skills; adding a new operation requires demonstrations for the action policy and retraining of the execution monitor. Sample removal is represented in recovery training, so the reported recovery rates do not establish generalization to unseen disturbances. Remaining failures include analyzer-docking misalignment and placement in an incorrect plate.

BibTeX

@article{robomlt2026,
  title   = {{Robo-MLT}: Dual-System Robotic Manipulation
             for Long-Horizon Medical Laboratory Tasks},
  author  = {Anonymous},
  journal = {Under review},
  year    = {2026}
}