LimX Oli Robot: Humanoid & Chemo Platforms

• LimX Oli Robot is a dual-platform research suite featuring a humanoid robot with whole-body control and a chemo-robotic system for automated oil droplet evolution.
• The humanoid platform utilizes 31 PD-controlled joints with advanced state representation learning and reinforcement learning for precise, high-frequency motion.
• The chemo-robotic system automates high-throughput experiments with precise liquid handling, computer vision-based droplet analysis, and robust workflow orchestration.

The LimX Oli robot designates two distinct yet technologically advanced robotic platforms: (1) a full-size humanoid robot equipped for high-dimensional whole-body control and reinforcement learning with advanced state representation learning (SRL) techniques and (2) a chemo-robotic liquid-handling system for automating the embodied chemical evolution of oil droplets. Both systems are frequently referenced as high-throughput, modular research platforms with precise actuation, rich sensor suites, and sophisticated software architectures for data-efficient experimental workflows (Yuan et al., 15 Dec 2025, Gutierrez et al., 2014).

1. Hardware Architectures

The LimX Oli humanoid robot comprises 31 actuated joints, each driven by high-torque brushless DC servo motors with embedded joint encoders. Motors operate via position commands within a proportional-derivative (PD) loop executed by onboard motor drivers. The full sensor suite includes per-joint positional ($q_j$) and velocity ($\dot q_j$) encoders, a 3-axis IMU (yielding angular velocity $\omega_t$ and gravity direction estimate $g_t$), and simulator-only privileged sensors for full body state, contact events, and terrain features. The controller runs on an industrial PC at 200 Hz and utilizes GPU-accelerated simulation platforms (Isaac Lab, MuJoCo XLA).

By contrast, the chemo-robotic LimX Oli (“DropBot”) platform consists of four physical subsystems: an XY gantry with NEMA 17/14 stepper motors and belt-driven axes for spatial positioning, servo-actuated syringes for liquid handling, a dispensing workstation with 96-well mixing plates and magnetic stirrers, and an imaging station equipped with a PS3 Eye camera (640×480 px, 30 FPS) beneath the glass stage. These modules are controlled by dual Arduino Mega microcontrollers (running modified Sprinter and custom firmware) and host Python modules for experimental logic (Gutierrez et al., 2014).

2. Control and Observation Spaces

Humanoid Robot Control

For whole-body control (WBC), the policy $\pi_\theta$ maps current observations $\bm o_t$ and high-level commands $\bm c_t$ to actions $\bm a_t \in \mathbb{R}^{31}$: $\bm a_t = \pi_\theta(\bm o_t, \bm c_t)$ Target joint positions are nominal plus action offsets; torques are computed as: $$\tau_t = K_p(q_\text{target} - q_t) + K_d(\dot q_\text{target} - \dot q_t)$$

Observation spaces include:

• Proprioceptive ($\bm o_t$): Directly sensed joint states (positions, velocities), IMU data, command velocities, and for mimic tasks, a reference pose.
• Privileged ($\bm s_t$): Full simulator state available to critic-only during training, including global kinematics, link velocities, contact indicators, and terrain features. By construction, $\bm o_t \subset \bm s_t$.

Chemo-Robotic Platform

The DropBot executes automated experiments via:

• Motion scheduling (XY carriage positioning)
• Reagent mixing and dispensing (servo-actuated plus pump-driven syringes)
• Droplet deposition at programmed coordinates
• Real-time imaging and computer vision for behavioral quantification (area, centroid, division, motion trajectories)
• Cleaning cycles and solution delivery based on G-code step delay parameters

3. Software Architecture and Real-Time Control

Humanoid Platform

Real robot control is implemented at high frequency (200 Hz) on an industrial PC; simulation/training is conducted on a single GPU using Isaac Lab and MuJoCo XLA. The policy and value networks utilize modular SRL frameworks, notably SRL4Humanoid, which permits high-quality implementations of multiple SRL methods (Yuan et al., 15 Dec 2025). Rollout data is processed in batches, and policy optimization leverages PPO integrated with PvP loss.

Chemo-Robotic Platform

Workflow orchestration is achieved via host-driven Python modules (RobotCtl API, PrintRun for G-code, genetic algorithm planning, and OpenCV-based computer vision). Motion control and pump interfaces rely on firmware running on two Arduino Mega controllers, enabling coordinated XY movement and fluidic operations. Real-time computer vision includes background subtraction (Gaussian mixture, OpenCV MOG, $\alpha=0.05$), edge/contour segmentation, watershed droplet separation, and behavioral feature extraction. G-code controls dispensing volumes, pump actuation delays, and syringe movements.

4. Learning Frameworks and Experimental Protocols

PvP Contrastive Learning (Humanoid)

PvP (Proprioceptive-Privileged) contrastive learning exploits the complementarity between proprioceptive and privileged state encodings. Training forms masked data pairs: $\tilde{\bm s}_t = \mathrm{ZeroMask}(\bm s_t)$ Encoders ($f_\theta$) and predictors ($h_\psi$) generate latent representations: $$\bm z = f_\theta(\bm s_t),\quad \tilde{\bm z} = f_\theta(\tilde{\bm s}_t)$$ With SimSiam-style negative cosine similarity: $$D_{\rm ncs}(\bm p,\,\bm q) = -\frac{\bm p}{\|\bm p\|_2}\cdot \frac{\bm q}{\|\bm q\|_2}$$ Total objective combines PPO and PvP loss: $L_\text{total} = L_\text{PPO} + \lambda\,\mathbbm{1}(t) L_\text{PvP}$ where $\lambda=0.5$, $\mathbbm{1}(t)=1$ every 50 steps.

Chemical Recipe and Droplet Protocols (DropBot)

Aqueous phase consists of 20 mM TTAB (pH=13), with oil mixtures formulated from four components: 1-octanol, 1-pentanol, diethyl phthalate, octanoic acid or dodecane, plus Sudan III dye. Automated deposition involves aspirating 80 µL mixed oil and depositing four 5 µL droplets at fixed coordinates. Cleaning cycles leverage acetone and aqueous flushes via programmable pump delays (flow rate ≈ 20 µL/s).

5. Experimental Benchmarks, Metrics, and Real-World Deployment

Humanoid Robot Results

• Sample efficiency: PvP attains velocity tracking reward (~1100) at 1×10^7 steps, outperforming vanilla PPO (requiring ~2×10^7). Mimic reward converges at ~1500 for PvP, significantly higher than PPO+SimSiam (~1350), PPO+SPR (~1300), PPO+VAE (~1100) at 3×10^7 steps.
• KPIs: PvP achieves 3× faster reduction in action smoothness penalty and ~15% lower joint-position RMSE in mimic compared to baselines.
• Ablations: Optimal SRL update interval is 50 steps; full batch for SRL gives maximal reward, though 50% batch realizes ~90% benefit; PvP loss must target policy encoder.
• Hardware performance: Sim2Sim transfer to MuJoCo yields <10% performance loss; controllers on LimX Oli hardware exhibit smooth, real-time walking and motion imitation (Yuan et al., 15 Dec 2025).

Chemo-Robotic Platform Results

• Throughput: One experiment runs in ~2 min (60 s recording, 60 s mixing/wash), enabling up to 30 experiments/h and 120 droplets/h. Continuous overnight operation (96 experiments) is feasible.
• Reliability: >95% droplet formation success across >2000 runs; main failure modes are needle clogging and incomplete cleaning.
• Behavioral metrics: Average droplet speed ~0.5 px/frame (~0.02 mm/frame), oscillation frequency 2–5 Hz, division yielding 2–4 droplets after 60 s in optimal recipes.
• Maintenance: Weekly full cleaning of tubing and pump syringes recommended (Gutierrez et al., 2014).

6. Practical Guidelines and Implementation Insights

• Proprioceptive and privileged states should be jointly leveraged in contrastive SRL for richer, augmentation-free representations.
• PvP loss is best applied intermittently (every 50 steps), targeting the policy encoder. Value network application may destabilize learning.
• PvP loss weight $\lambda$ of 0.5 yields stable PPO optimization; weights ≥1.0 risk instability.
• Maximal SRL benefit is obtained from utilizing the full rollout batch. Partial batches yield near-maximal gains for mimic tasks.
• Latent representations learned via PvP encode task-relevant features, integrating seamlessly into PPO policy heads for 2×–3× improved convergence.
• Chemo-robotic experimental parameters (reagent selection, flow rates, deposition protocols) are optimized for high success rates and minimal failure modes, under rigorous safety constraints and reproducible hardware/software architectures.

7. System Integration and Extensibility

Both LimX Oli platforms support modular expansion. For the humanoid robot, SRL4Humanoid allows plug-and-play evaluation of alternative SRL algorithms under unified benchmarking conditions. The DropBot supplies open-source hardware (3D printed components, full bill of materials), Arduino-compatible firmware, and host Python code, facilitating replication and experimental extension. Supplementary repositories contain CAD files (3DPP1–16), G-code scripts, and protocol documentation (Yuan et al., 15 Dec 2025, Gutierrez et al., 2014).

These system-level design choices make LimX Oli an archetypal model for reproducible, data-efficient robotic and chemo-robotic experimentation in both computational and physical sciences.