ErgoCub Humanoid Robot

Updated 6 October 2025

ErgoCub is an advanced humanoid robot designed for ergonomic human–robot interaction, featuring optimized hardware, sensor integration, and partner-aware control mechanisms.
It utilizes bilevel and model-based optimization to reduce energy expenditure by up to 43% and lower critical joint torques during payload handling.
The platform integrates multimodal sensor fusion and bio-inspired motion generation to achieve natural locomotion, safe obstacle avoidance, and dynamic adaptability in collaborative tasks.

ErgoCub is an advanced humanoid robot platform developed to optimize ergonomic interaction with humans across a spectrum of collaborative physical tasks. Distinguished by its co-designed mechanics, control algorithms, and sensor architecture, ErgoCub's development amalgamates model-based optimization, partner-aware control strategies, sensor fusion techniques, and bio-inspired motion generation. These integrated capabilities enable real-time posture adaptation, safe whole-body interaction, ergonomic payload lifting, and natural locomotion patterns, with strong empirical validation supported across recent experimental and theoretical work.

1. Hardware Design, Dynamics, and Parametric Optimization

The physical parameters of ErgoCub—link geometries, densities, and motor specifications—are optimized to enhance ergonomic collaboration during payload lifting and related tasks (Sartore et al., 2022, Sartore et al., 2023). Each rigid link is parameterized via multipliers on length and density; their mass, center of mass, and inertia matrix are expressed as integrals over the associated volume, e.g.:

$m(l, \rho) = \iiint_{V(l)} \rho(r) dr, \quad I_a(l, \rho) = -\iiint_{V(l)} \rho(r) [S(r)]^2 dr, \quad c(l, \rho) = \frac{\iiint_{V(l)} r \rho(r) dr}{m(l, \rho)},$

where $l$ encodes geometry, $\rho$ is density, and $S(r)$ is the skew-symmetric matrix for $r$ .

For the overall multibody system, dynamics are augmented by hardware parameters $\pi$ , yielding:

$M(q, \pi) \dot{\nu} + h(q, \nu, \pi) = B\tau + J_c^T(q,\pi) f,$

with $q$ the floating-base and joint positions, $\nu$ velocities, $\tau$ actuator torques, and $f$ external/contact forces.

The hardware optimization problem, cast as a nonlinear program (using solvers such as CasADi and interior-point methods (Sartore et al., 2022)), seeks joint optimality over:

Robot energy expenditure ( $||\tau(q, \pi)||_2^2$ ),
Foot/posture stability,
Center-of-mass height (robustness),
Link densities (physical feasibility), and incorporates human/robot coupled dynamics to preserve human ergonomics during interaction.

A bilevel optimization approach (Sartore et al., 2023) uses a genetic algorithm at the outer loop to select design candidates $\pi$ , while the inner loop evaluates static and dynamic ergonomic indexes using nonlinear optimization:

$\bar{M}(q,\pi)\dot{\nu} + h(q,\nu,\pi) = \bar{B}(\pi)\bar{\tau} + J_c^T(q,\pi) f - \bar{K}_v(\pi)\nu,$

where motor dynamics and viscous friction terms ( $\bar{K}_v(\pi)$ ) are part of the robot model.

Key outcomes:

The optimized hardware design extends load lifting heights (0.8–1.5 m for ergoCub, vs 0.8–1.2 m for iCub) (Sartore et al., 2022).
Robot energy expenditure is reduced up to 43% over pure nonlinear-optimized designs in static tasks (38% in dynamic actions), with critical human joint torques (e.g. L5S1) reduced up to 30% (Sartore et al., 2023).
The framework generalizes well across multi-human, multi-load scenarios.

2. Sensing Architecture and Real-Time Sensor Fusion

ErgoCub utilizes distributed, multimodal sensors and advanced fusion algorithms for contact-aware control (Sorrentino et al., 28 Feb 2024). Sensor modalities include:

Joint encoders (kinematic data),
Motor current sensors (indirect torque estimation),
Force/Torque (FT) sensors (wrench measurement),
Inertial sensors (accelerometer, gyroscope).

The core fusion algorithm is based on the Unscented Kalman Filter (UKF), allowing estimation of joint torques without direct sensors:

$x_k = [\hat{\dot{s}}^T, \hat{\tau}_m^T, \hat{\tau}_F^T, \hat{f}_{FT}^T, \hat{f}_{ext}^T]^T$

$x_{k+1} = F(x_k, u_k, v_k)$

$y_k = G(x_k, u_k, w_k)$

with $u_k$ including robot configuration, process noise $v_k$ , and measurement noise $w_k$ .

Experiments on leg and torso joints show that UKF-based torque estimation achieves lower RMSE in tracking (down to 0.08–2.5 Nm), outperforming recursive Newton-Euler algorithms especially in scenarios involving external contacts. The fusion of distributed sensors ensures resilience to non-ideal conditions and facilitates compliant human–robot interaction.

3. Control Architectures for Human–Robot Interaction

Partner-aware control is central to ErgoCub’s collaborative performance (Tirupachuri et al., 2018, Rapetti et al., 2021). Rather than treating the human as an uncertainty, system dynamics are formulated in a coupled formalism:

$M_{tot}(q_R, q_P)\,\ddot{q}_{tot} + h_{tot}(q_R, q_P, \dot{q}_R, \dot{q}_P) = \tau_{tot} + J_{tot}^T f_{int}$

where $q_R, q_P$ are robot and partner (human or robot) configurations, and $f_{int}$ is the interaction force.

Task-level controllers solve optimization problems such as:

$\tau_R^\ast = \arg\min_{\tau_R} \|e_{task}\|^2 + \lambda\|\tau_R - K_{partner} f_{int}\|^2$

which adjust effort based on partner assistance, ensuring energy efficiency and stability.

In multi-agent scenarios, a centralized controller (Rapetti et al., 2021) coordinates momentum and force distribution:

Tracks individual and system momentum,
Optimizes contact/grasp forces:

$f^\ast = \arg\min_f \|K_\tau \tau(f)\|_2$

subj. to friction cone, momentum and pose constraints,

Schedules postural trajectories offline to minimize joint torque norms, ensuring smooth transitions (minimum jerk profiles).

Experiments with two iCub robots demonstrate minimized torque, accurate payload tracking, and adaptive load sharing—features being directly ported to ErgoCub’s architecture for human–robot and robot–robot collaboration.

4. Ergonomic Human–Robot Interaction Protocols

Real-time ergonomic optimization is implemented via continuous observation and correction of the human operator’s posture (Shafti et al., 2018). Sensor inputs (RGB-D camera, inertial measurement units) are calibrated against a neutral "reference posture," e.g.:

$R_{ref} = R_{wrist}^{-1} R_{lower-arm}$

Joint angles for ergonomic assessment (Rapid Upper Limb Assessment, RULA) are calculated from skeleton vectors, for example:

$\theta_{sag} = \cos^{-1}\left(\frac{(r_h - r_e)^\top (r_e - r_s)}{\|r_h - r_e\|\, \|r_e - r_s\|}\right)$

When deviation from ergonomic thresholds is detected, the algorithm identifies six canonical causes and applies corresponding robot-supported workpiece manipulation:

Cause	Response	Description
Upper-arm sagittal deviation	Workpiece translation (forward/back)	Re-align to ergonomic window
Upper-arm abduction	Lateral workpiece shift	Bring arm to vertical
Lower-arm sagittal deviation	Vertical workpiece translation	Target lower-arm angle of ~90°
Lower-arm transversal	Reposition workpiece sideways	Correct alignment
Wrist transversal deviation	Rotate workpiece	Realign wrist
Wrist sagittal deviation	Vertical workpiece translation	Achieve wrist neutrality

Robotic responses apply translation/rotation based on individually calibrated limb lengths, using explicit vector formulas. Experimental protocols—replicating those originally run on the Baxter Research Robot—are adapted for ErgoCub to quantify ergonomic improvement via RULA and EMG scores during both human-only and robot-assisted conditions.

5. Bio-inspired Reactive Motion Generation and Multimodal Perception

ErgoCub's whole-body motion controller leverages quadratic programming (QP) for redundant DoF management, constraint satisfaction, and visuo-tactile-aware obstacle avoidance (Rozlivek et al., 2023). The controller simultaneously commands both 7-DoF arms and a 3-DoF torso (17 DoF):

$\nu = J(q)\,\dot{q}, \quad \nu - \lambda = J(q)\,\dot{q}$

Cost function for QP:

$\min\ \frac{\mu}{2} \dot{q}^T W_q \dot{q} + \frac{1}{2} \lambda^T W_\lambda \lambda + \frac{c_h}{2} (\dot{q} - \dot{q}_h)^T W_q (\dot{q} - \dot{q}_h)$

with manipulability index damping $\mu$ .

Obstacle avoidance constraints are unified across modalities (visual, proximity, tactile):

$n_C^T (J_C \dot{q}) < x_a$

$x_a = (k_1 - V_C a_t) \cdot k_2$

where $a_t$ is threat level and $V_C$ sensor gain.

Trajectory waypoints are generated to follow minimum-jerk profiles via an optimized LTI system, and orientation is interpolated using Slerp for human-like movement.

Validation includes simulation (high reachability, smooth self-collision avoidance) and real-world tasks, including interactive HRI scenarios (e.g., tabletop board games) where the controller safely reacts to proximity and contact in dynamic environments.

6. Locomotion: DNN-Driven MPC and Physics-Informed Stabilization

ErgoCub's locomotion architecture integrates an autoregressive Deep Neural Network trained on human MoCap data, ensuring stylistic gait patterns (Romualdi et al., 10 Oct 2024, D'Elia et al., 29 Sep 2025). The pipeline features:

(a) Deep Neural Network Layer

Generates centroidal and postural reference trajectories, regularizing the motion planning process.
Training minimizes:

$\text{Loss} = \|\text{Reference trajectory} - \text{Predicted trajectory}\|^2$

with AdamWR optimizer, 110 epochs.

(b) Trajectory Adjustment Layer

Direct multiple shooting optimization for COM tracking and step adjustment.
Key tasks:

$\Psi_h = \|h^n - h\|^2_{\Lambda_h} + \|p^n_{\text{CoM}} - p_{\text{CoM}}\|^2_{\Lambda_{\text{CoM}}}$

Control Barrier Functions (CBF) enforce safe COM heights, e.g.:

$g(t) = -\alpha (p_{\text{CoM},z}^{\min} - p_{\text{CoM},z})(p_{\text{CoM},z}^{\max} - p_{\text{CoM},z})$

RHP and MPC variants: MPC uses Kalman-filtered velocity estimates with filter parameters tuned via genetic algorithms.

(c) Trajectory Control Layer

COM–ZMP controller for dynamic feasibility:

$\dot{x}^*_{G} = \dot{x}^\text{ref}_{G} - K_\text{ZMP}(x^\text{ref}_\text{ZMP} - x_\text{ZMP}) + K_G (p^\text{ref}_G - p_G)$

QP-based inverse kinematics produces smooth joint commands for foot and posture trajectories.

During network training, a physics-informed loss term $\mathcal{L}_B$ enforces foot contact stability:

$v_{set} = J_{SF}\, \nu = 0$

PI correction block during inference adds proportional-integral feedback to the predicted state:

${}^{\mathcal{B}}\omega^c_{\mathcal{I},\mathcal{B}} = {}^{\mathcal{B}}\hat{\omega}_{\mathcal{I},\mathcal{B}} - k_1\, \text{skew}(({}^{\mathcal{I}}R^d_{\mathcal{B}})^T {}^{\mathcal{I}}R_{\mathcal{B}})^\vee$

This dual strategy reduces trajectory drift and foot-sliding, with observed improvements in RMSE for foot velocities, contact timing, and overall locomotion stability.

Experimental results show recovery from impulsive disturbances up to 68 N, natural walking styles, and stable trajectory following, with resource materials provided for replication and further analysis (Romualdi et al., 10 Oct 2024).

7. Impact and Significance in Human-Robot Ergonomic Collaboration

The ErgoCub platform demonstrates the application of model-based design, sensor fusion, partner-aware control, ergonomic optimization, and bio-inspired motion strategies. Measured outcomes include:

Expanded load-lifting range vs predecessor robots (iCub),
Significant reductions in energy expenditure and internal joint torque,
Enhanced safety, compliance, and responsiveness in physical human–robot interaction,
Human-like motion generation under dynamic, contact-rich scenarios,
Validation through quantitative metrics (RULA, EMG, RMSE) across static and dynamic collaborative tasks.

The system’s co-design methodologies—bilevel optimization, dynamic sensor fusion, unified control frameworks—set a reference for ergonomic humanoid robot development suitable for multi-agent industrial, healthcare, and service environments. A plausible implication is that such integrated architectures could underpin the next generation of collaborative robots with inherently safer and more energy-efficient interaction.