RubiCap: Capping & RL Framework

• RubiCap is a dual-purpose system combining a compact, open-source lab capping machine for confined spaces with a rubric-guided reinforcement learning framework for dense image captioning.
• The mechanical capping device features precise actuation, vision-based fault detection, and modular design to deliver reliable, cost-effective sample handling in self-driving laboratories.
• The RL framework leverages LLM-generated rubrics as structured rewards, achieving superior dense captioning performance and maintaining generalization without catastrophic forgetting.

RubiCap refers to two independent systems at the intersection of robotics and artificial intelligence: (1) an open-source compact capping machine for confined laboratory spaces, and (2) a rubric-guided reinforcement learning framework for dense image captioning. Both systems address critical automation challenges by introducing open design, performance-driven engineering, and robust evaluation mechanisms. The following sections provide a detailed exposition of both instantiations, each grounded in peer-reviewed literature and empirical benchmarks.

1. RubiCap as an Open-Source Confined-Space Capping Machine

RubiCap is an open-source, automation-ready capping and uncapping machine specifically engineered for confined laboratory environments such as fume hoods (Munguia-Galeano et al., 4 Jun 2025). The system is designed to deliver high reliability for self-driving laboratories (SDLs) where automated sample handling is essential but commercial solutions are prohibitively bulky and often closed-source.

Key Mechanical and Electrical Features

The device maintains a compact 0.06 m³ footprint (300 mm × 500 mm × 400 mm), compatible with standard fume hood dimensions. The frame utilizes 20 × 20 mm anodized aluminum extrusion profiles and T-slot connectors for structural rigidity and ease of assembly. Actuation is driven by a 12 V DC, 19.8 W, 59 N·cm stall torque brushed, geared motor coupled via a T5 timing belt (10 mm width, 340 mm length, 68 teeth, 1:1 ratio) to a 10 mm flanged lead screw. The lead screw-mechanized vial holder employs 3D-printed resin, an Igus WW-10-40-10 carriage, WS-10-40-600 linear rail, twin guide rods, and a parallel cam-based locking linkage for robust sample engagement.

Sample handling is gravity-fed: caps drop into the capping lane as vials pass below, where a rubber-lined, spinning profile applies torque by friction for capping or, in reverse, uncapping.

Power is supplied via dual DRL-30 switch-mode supplies (85–264 V AC to 12 V DC, 2.1 A). Motor direction is controlled through an H-bridge implemented with four LY2-DC12 relays, coordinated by an Arduino Uno microcontroller interfaced with SRD-05VDC-SL-C relays and opto-isolated for electrical safety.

Manual operation is possible through a spring-return push-button panel with "Home," "Cap," and "E-Stop" functions, alongside full shutdown safety mechanisms.

2. Vision-Based Fault Detection and Control System

RubiCap’s integrated vision system differentiates it from typical industrial capping solutions by providing automated, in-line error detection (Munguia-Galeano et al., 4 Jun 2025).

A standard USB RGB camera is positioned ~15 cm from the capping lane with uniform LED backlighting. The algorithm, implemented in OpenCV/Python, processes each vial after actuation using the following pipeline:

1. Extract region of interest (ROI) around vial threads.
2. Convert ROI to HSV, apply a blue-channel mask to isolate reflections through unsealed threads.
3. Employ grayscale thresholding and contour detection to compute the total "bright" area $A$.
4. If $A > A_{thresh}$, the vial is flagged "Not Capped" (red); otherwise, "Capped" (green).

The Arduino reads the camera verdict via a digital input. An uncapped detection triggers an immediate stop and alert, ensuring no failed samples proceed undetected.

Control logic is encoded as a state machine operating through Idle → Home → Load Vial → Cap → Vision Check → Uncap → Vision Check → Return Vial → Idle, with embedded emergency-stop and manual-override options at all stages.

3. Experimental Validation and Quantitative Benchmarking

RubiCap’s reliability and sealing integrity were validated under rigorous protocols (Munguia-Galeano et al., 4 Jun 2025):

• Robustness: In 100 consecutive automated cycles (pick–place–cap–uncap) using a Franka Panda robot in a fume hood, RubiCap achieved a 100% success rate.
• Sealing Capacity: Three solvents (distilled water, ethanol, acetone, four replicates each) were capped by RubiCap, a Chemspeed industrial station, and manually. Mass loss was logged gravimetrically at 3 h intervals over 72 h.

Method	Water Loss/day	Ethanol Loss/day	Acetone Loss/day	Average Loss/day
RubiCap	0.0322%	0.2036%	1.3956%	0.54%
Manual	0.0062%	0.0071%	0.0152%	0.013%
Chemspeed	0.0041%	0.0053%	0.0141%	0.0078%

Chemspeed and manual methods exhibit superior sealing for high-volatility solvents, attributed to closed-loop torque control absent in RubiCap. Crucially, only RubiCap integrates vision-based failure detection; Chemspeed offers no cap verification, while manual capping relies on operator judgment.

Throughput for RubiCap is single-lane (~10–15 s per cycle, robot-limited), with a deployable footprint and cost substantially below industrial (£50k+) competitors. Bill-of-materials costs were: aluminum profiles (<£100), DC motor (<£100), Arduino (<£30), 3D-printed parts (<£50), camera (<£40).

4. Integration, Adaptability, and Practical Deployment

RubiCap’s modularity enables rapid adaptation for various laboratory requirements (Munguia-Galeano et al., 4 Jun 2025). Its robot-agnostic interface (standard pick-and-place), simple serial/TTL communication, and modular expansion accommodate additional capping lanes, vial sizes, or torque-feedback options with minimal redesign.

Open-source mechanical drawings, firmware, vision code, and bill of materials are released under a permissive license, facilitating local fabrication and integration into any SDL workflow. The design ensures minimal bench footprint and compatibility with standard fume-hood sashes, featuring thorough cable protection, DIN-rail terminal blocks, and extensive fail-safes (E-stop, minimal pinch points).

5. RubiCap as a Rubric-Guided Reinforcement Learning Framework for Dense Image Captioning

Distinct from the mechanical context, RubiCap also denotes a rubric-guided reinforcement learning (RL) method addressing the verification bottleneck in dense image captioning (Huang et al., 10 Mar 2026). Dense captioning requires region-level, fine-grained descriptions beyond global image summaries, directly impacting vision–LLM pretraining and text-to-image generation.

Supervised fine-tuning (SFT) with synthetic captions from teacher VLMs suffers from linguistic collapse, catastrophic forgetting, and distributional mismatch. Standard RL with n-gram rewards or holistic VLM-judge scores is unsuited to open-ended, subjective captioning, leading to degenerate reward hacking.

RubiCap introduces automated, sample-specific LLM-written rubrics as structured rewards within an RL framework. For each image, a committee of K diverse VLMs generates candidate captions. An LLM rubric writer extracts consensus elements, diagnoses deficiencies in the student policy’s output, and formulates M binary pass/fail criteria $\mathcal{R}(x) = \{(r_m, w_m)\}_{m=1}^M$, each with severity weight $w_m \in \{1.0, 2.0, 3.0\}$.

At each RL step, multiple student rollouts are judged per criterion by an LLM, and reward is computed as:

$$G_i = \frac{\sum_{m=1}^M w_m \hat{y}_{i,m}}{\sum_{m=1}^M w_m}$$

Relative advantages $A_i$ are computed within the batch, and policy parameters updated by a clipped-PPO objective (Group Relative Policy Optimization, GRPO). Deficiencies are classified as Critical ($3.0$), Important ($2.0$), or Minor ($1.0$), depending on the error type.

6. Empirical Evaluation and Outcomes in Dense Captioning

RubiCap was evaluated on key dense captioning and vision-language pretraining benchmarks (Huang et al., 10 Mar 2026):

• Datasets: PixMoCap and DenseFusion-4V-100K, each with 50,000 expert or GPT-4V-augmented images.
• Baselines: Untuned student model, SFT on synthetic data, RL with ROUGE-L rewards, VLM-judge RL, CapRL-3B.
• Benchmarks: CapArena (pairwise win rate), CaptionQA (QA accuracy under word constraints), VLM generalization, expert blind ranking, pretraining utility for downstream VLMs.

Salient results include:

• RubiCap-7B achieved a +20.8% win rate on CapArena over base (70.8% vs 50%), outperforming all baselines.
• On PixMoCap, RubiCap-7B outperformed expert ground-truth captions with a 66.0% win rate.
• Word efficiency gains: RubiCap-7B matched 32B-scale Qwen2.5-VL-32B at ~7B parameter cost.
• Pretraining on RubiCap-3B annotations led to +3.4% higher downstream task accuracy than pretraining on GPT-4V captions.
• RubiCap models avoided catastrophic forgetting, maintaining generalization on diverse vision-language tasks, in direct contrast to SFT methods.

7. Limitations, Current Challenges, and Prospects

Both instantiations of RubiCap exhibit context-specific limitations:

• Capping machine: Sealing integrity is suboptimal for high-volatility solvents compared to industrial torque-controlled cappers. The vision system’s fault detection depends on correct algorithmic thresholds and camera alignment. Adaptation of torque feedback is an identified pathway for enhancement (Munguia-Galeano et al., 4 Jun 2025).
• RL framework: Rubric quality is dependent on LLM writer fidelity and committee diversity. Generating sample-specific rubrics incurs up-front cost. Weighting of rubric criteria is manually tuned; adaptive learning of reward structures is a suggested direction for future work.
• Scalability: RubiCap’s machine is single-lane but modular; the RL variant hinges on VLM and LLM compute availability.
• Extensibility: RubiCap RL has the potential to extend to other open-ended tasks—including summarization and code generation—by rubric generalization.

These systems advance automation in both laboratory operations and AI model development by embedding open-source ethos, robust quality assurance, and domain-aware reward shaping at their core (Munguia-Galeano et al., 4 Jun 2025, Huang et al., 10 Mar 2026).