Pandora: Structurally Elastic Humanoid Robot

• PANDORA is an open-source humanoid platform featuring distributed structural elasticity achieved through 3D-printed compliant load-bearing links.
• It integrates hybrid additive-manufacturing with precise metal parts to lower part count and streamline assembly, cutting manufacturing time and cost.
• Experimental validations demonstrate robust balancing and stepping despite challenges in state estimation and joint control due to inherent structural compliance.

PANDORA is an open-source humanoid robot developed at Virginia Tech and presented as a structurally elastic research platform for legged and whole-body humanoid control. Its defining feature is that a majority of its structural elements are manufactured using 3D-printed compliant materials, so that compliance is distributed through the robot’s load-bearing links rather than localized in actuator springs. In contrast to conventional series elastic actuators or flexible-joint mechanisms, PANDORA is designed to be compliant under load at the structural level, a design choice that lowers manufacturing cost and time, design complexity, and assembly time while introducing substantial challenges in state estimation, joint control, and whole-body control (Herron et al., 2024).

1. Definition and design premise

PANDORA’s central mechanical idea is what the authors call “structural elasticity.” Instead of adding springs inside the transmission, the printed structural members themselves are designed to be compliant under load. This means that the robot is effectively elastic at the link level. The intended consequence is a substitution of structural compliance for the compliance traditionally achieved through elastic actuators, with corresponding reductions in part count, design complexity, cost, and manufacturing time (Herron et al., 2024).

This design differs fundamentally from standard series elastic actuator and flexible-joint humanoid architectures. In those systems, compliance is localized and intentionally modeled as part of the actuator mechanism, leaving the links closer to rigid and the elasticity represented by a known spring element. In PANDORA, compliance is distributed throughout the printed load-bearing links. The authors emphasize that this effectively introduces additional unmodeled degrees of freedom into the structure itself, so the rigid-body assumption underlying much of conventional humanoid estimation and control is violated (Herron et al., 2024).

The complete design is open-source and includes mechanical CAD, PCB and firmware, and high-level software, all made available through TREC’s GitLab. This positions PANDORA as a publicly reusable research platform rather than a one-off prototype. A plausible implication is that the robot is meant not only as a demonstrator of additive-manufactured humanoid mechanics, but also as an experimental substrate for studying estimation and control under distributed structural compliance (Herron et al., 2024).

2. Mechanical architecture and additive manufacturing

Mechanically, PANDORA is a hybrid additive-manufacturing/subtractive-manufacturing machine. Most structural components are 3D printed, while critical joints and actuators retain some subtractive-manufactured metal parts for precision and strength. The robot stands 1.9 m tall and weighs 49 kg (Herron et al., 2024).

The lower body illustrates the intended manufacturing economy. It was reduced to 228 parts, compared with 510 parts in THOR and 480 in ESCHER, which is more than a 50% reduction in part count. In the lower body, additive-manufactured parts account for 6.4 kg of 23.27 kg, or 27.5% of the mass. Assembly time for the lower body is reported as 8 hours with two people (Herron et al., 2024).

The main lower-body structural elements are the pelvis, thigh, and shin, printed on a Creality CR-10. These parts were shaped to maximize strength and minimize weight, and they incorporate functional features including slide rails for sensor mounting and calibration slots for encoder zeroing. The upper body includes a chest that houses batteries, a mini PC, a network switch, and the shoulder structure; the head and neck support a Hokuyo LiDAR and a Carnegie Robotics Multisense S7; and the arms use a biomimetic mix of direct-drive and cable-driven actuation to keep mass close to the body (Herron et al., 2024).

The paper emphasizes that structural compliance in 3D-printed members depends on geometry, material, print orientation, print quality, and settings. It notes that printing perpendicular to expected load directions is an important practical rule, but that accurate pre-print prediction remains difficult. This makes PANDORA’s elasticity mechanically simple in one sense—because it removes dedicated spring elements—but more difficult to model predictively than actuator-localized compliance (Herron et al., 2024).

3. Electrical system, sensing, and software stack

PANDORA uses an in-house modular electrical architecture built around Low-Level Controllers, or LLCs. Each LLC consists of an EasyCAT shield, a TIVA microcontroller, and a Sensor Interface shield, and each can manage two actuators and two joints. The LLC collects absolute joint encoders, motor quadrature encoders, motor current, and force sensor data at 1000 Hz. It also reads ground reaction force/torque at the ankles using ATI FT sensors (Herron et al., 2024).

Sensor signals are conditioned with hardware filtering to remove 60 Hz mains hum and motor noise. The LLC sends PWM commands to the Motor Control Unit and runs a force-impedance controller based on a disturbance-observer approach. At the system level, the robot’s state is sent to the high-level controller at 500 Hz over EtherCAT using an EasyCAT Pro shield, with the six lower-body LLCs arranged in a daisy-chained master-slave network. The high-level computer acts as the EtherCAT master (Herron et al., 2024).

The software stack uses TREC Robotics Software (TRS), which interfaces with IHMC’s Open Robotics Software (ORS). ORS provides the footstep planner, DCM-based center-of-mass planner, and whole-body controller. Safety logic is built into each LLC, including encoder and force limit checks that can trigger a HALT state and power down all motors. The paper also notes that real-time operating system support is used on the TIVA microcontroller to enforce timing constraints (Herron et al., 2024).

The combination of distributed sensing, 1000 Hz local control, and 500 Hz high-level EtherCAT communication reflects the paper’s broader thesis: PANDORA’s hardware is mechanically simplified by structural elasticity, but the informational and control burden shifts toward estimation, synchronization, and compensation for compliance-induced mismatch.

4. Joint actuation, structural compliance, and estimation

PANDORA’s actuator-joint structure is not a simple one-to-one chain. The hip roll/yaw and ankle roll/pitch joints are coupled so that both actuators contribute to both joints, while hip and knee pitch are driven more directly. As a result, accurate joint-angle mapping is critical (Herron et al., 2024).

At the joint level, actuator forces are computed from desired joint torques using an inverse Jacobian-transpose mapping: $$\mathbf{f}_{\rm act} = \mathbf{J}^{\rm -T}(\mathbf{q}_{\rm joint})\boldsymbol{\tau}_{\rm joint}$$ where $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$ is the actuator force, $\mathbf{q}_{\rm joint} \in \mathbb{R}^{2}$ are the two joint angles for a coupled joint pair, $$\mathbf{J}(\mathbf{q}_{\rm joint}) \in \mathbb{R}^{2\times 2}$$ is the mechanism Jacobian, and $\boldsymbol{\tau}_{\rm joint} \in \mathbb{R}^2$ are the desired joint torques (Herron et al., 2024).

The kinematic relation used for actuator and joint position estimates is

$$\mathbf{q}_{\rm act} = f_{j,a}(\mathbf{q}_{\rm joint}) = f_{m,a}(\mathbf{q}_{\rm motor})$$

where $\mathbf{q}_{\rm act} \in \mathbb{R}^{2}$ are actuator positions, $f_{j,a}$ maps joint space to actuator space, and $f_{m,a}$ maps motor encoder measurements to actuator space. In the ideal no-load case, these mappings agree, but with compliance and backlash they can drift out of phase, causing torque-tracking errors and instability (Herron et al., 2024).

The paper states that a flexible-joint model might improve stability, but does not use a full-order compliant model in the controller. Instead, the joint-level controller uses a disturbance-observer-based force loop with an ideal titanium leaf-spring actuator model as the nominal reference, even though the actual PANDORA implementation removes those leaf springs and relies on printed structural compliance. The disturbance observer compensates for model mismatch, helps remove stiction, and can help deal with backlash, but it can also excite backlash and lead to chattering or even instability (Herron et al., 2024).

To mitigate this, the controller introduces a disturbance-observer gain,

$0 \leq k_{dob} \leq 1$

with reported values of $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$0 for the hip and ankle actuators and $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$1 for the thigh/knee actuators. These values reflect a tradeoff between disturbance rejection and backlash-induced oscillation (Herron et al., 2024).

5. Whole-body control and state-estimation challenges

Structural compliance affects whole-body control primarily through state estimation. Since the whole-body controller assumes a largely rigid kinematic model, link deflection creates errors in estimated foot pose, base orientation, and center-of-mass location. The paper shows an example where the right foot deflects by up to 2 cm when the pelvis is manually perturbed, making the virtual robot’s kinematics visibly different from the real robot’s physical state. During contact transitions, these errors can become even larger (Herron et al., 2024).

The authors note that if the estimator believes the feet have moved when they have not, the whole-body controller can generate incorrect momentum commands and potentially destabilize the robot. They mention a heuristic from prior work that adds stiffness based on applied torque, but because PANDORA does not directly measure joint torques, they are exploring a Kalman filter that fuses motor encoders, joint encoders, and force sensors to improve joint position estimates. This suggests that the principal unresolved problem in the platform is not gross actuation capacity but compliant-state observability (Herron et al., 2024).

In the balancing experiments, the whole-body controller uses eight tasks: both feet are held in task-space position with zero desired acceleration during stance, both feet have orientation control, the pelvis orientation is controlled, whole-body linear momentum is regulated, and a privileged joint-space task biases the knees away from singular configurations. Positive ground reaction force and dynamic feasibility constraints are enforced (Herron et al., 2024).

This control design remains structurally conventional in the sense that it is still a task-based whole-body controller built on a rigid-body framework. The novelty lies not in replacing whole-body control theory, but in testing how far standard methods can be pushed when the underlying robot departs from rigid-link assumptions.

6. Experimental validation and performance

The paper reports two main hardware locomotion demonstrations. The first is robust double-support balancing under external disturbances. Over a 12-second trial, an operator pushes the pelvis in the $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$2, $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$3, $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$4, and $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$5 directions. During the perturbations, the center of mass is displaced in the $\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$6-$\mathbf{f}_{\rm act} \in \mathbb{R}^{2}$7 plane, but after the force is removed the controller returns it to equilibrium. Estimated foot positions drift significantly, with black traces over the foot polygons indicating up to 5 cm of apparent foot motion during the balance experiment. Despite this, the robot remains stable and recovers (Herron et al., 2024).

The second experiment is in-place stepping. The same controller transitions between double support and single support, with one foot moving through a swing trajectory while a DCM-based LQR controller drives the center of mass along a desired footstep plan. Because the robot is supposed to step in place, the footstep reference is repeatedly reset to the current stance positions. PANDORA can complete several steps in a row, specifically 5 contiguous footsteps, before losing balance (Herron et al., 2024).

Walking is not yet fully stable. The main failure modes are torque-tracking errors, momentum-estimation errors, and, most importantly, kinematic errors caused by structural elasticity during single support when the stance leg bears the full body load. The right-foot tracking behavior is not plotted, but is described as roughly mirroring the left foot (Herron et al., 2024).

These results establish that structurally elastic full-size humanoid hardware can support robust balancing and short stepping sequences under standard whole-body control methods, while also making clear that the dominant limitations arise from estimation and model mismatch rather than from basic mechanical integrity.

7. Position within humanoid robotics research

PANDORA is positioned as an open-source humanoid platform that combines three features that are uncommon together: a full-size humanoid form factor, structural compliance achieved through printed load-bearing members, and a public software-hardware stack for whole-body control (Herron et al., 2024).

Its contribution is therefore not simply the introduction of another additive-manufactured robot. The paper argues that a robot can be built with compliance distributed through its structure rather than localized in elastic actuators, and that standard whole-body control methods can still produce useful behaviors such as balancing and short stepping sequences. At the same time, it makes clear that this shifts the principal bottleneck from manufacturing to estimation and control (Herron et al., 2024).

A common misconception would be to treat structural elasticity as equivalent to conventional series elasticity. The paper does not support that interpretation. In PANDORA, compliance is harder to model, depends on manufacturing variables such as geometry and print orientation, and introduces effectively unmodeled degrees of freedom in the links themselves. The benefit is simpler mechanical construction; the cost is a more difficult estimation and control problem (Herron et al., 2024).

For research on legged and whole-body humanoid control, PANDORA therefore functions as both a platform and a problem formulation. It demonstrates that structurally elastic humanoids are mechanically practical and experimentally viable, while motivating future work on better compliant-state estimation, particularly Kalman filter-based fusion of encoders and force sensing (Herron et al., 2024).