TALOS: Robotics Platform and Acronym
- TALOS is a full-size humanoid robot with 32 degrees of freedom that serves as an experimental platform for whole-body control, contact regulation, and physical human assistance.
- It couples high-level decision making with real-time multi-contact and force-aware execution using techniques like SEIKO retargeting and hierarchical flow matching at high control frequencies.
- Beyond robotics, TALOS also represents various domain-specific acronyms for methods in semantic scene completion, privacy-preserving learning, and spacecraft design.
TALOS is used in contemporary research in two distinct ways. Most prominently in robotics, it denotes the full-size humanoid robot developed by PAL Robotics and used as an experimental platform for whole-body control, multi-contact locomotion and manipulation, language-guided support-contact selection, and physical human assistance. In parallel, the same name—variously capitalized as TALOS, Talos, TALoS, or TaLoS—is also used as an acronym for unrelated methods and toolchains in semantic scene completion, privacy-preserving contrastive learning, graph adversarial defense, key-value-store replication, spacecraft design, autonomous experimental control, secure enclave migration, vulnerability mitigation, sparse model editing, and recommender-system optimization (Rouxel et al., 2023, Rouxel et al., 2024, Totsila et al., 2024, Razmjoo et al., 2023, Jang et al., 2024, He et al., 2021, Li et al., 2024, Vardoulakis et al., 2021, Gandarillas et al., 2023, Volponi et al., 2024, Vasileaidis et al., 5 Sep 2025, Huang et al., 2017, Iurada et al., 3 Apr 2025, Zhang et al., 27 Jan 2026).
1. TALOS as a humanoid research platform
In the robotics literature, TALOS is described as a full-size humanoid bipedal robot from PAL Robotics. Several papers give a consistent embodiment profile: height $1.75$ m, mass $99.7$ kg, and $32$ DoF, while specific experiments often control only a subset of joints depending on the task and controller design (Rouxel et al., 2024, Rouxel et al., 2023). The platform appears in real-robot studies on pushing, far-reaching, stair climbing, stepping on sloped surfaces, dishwasher closing, non-prehensile box pushing, support-contact placement, sit-to-stand assistance, and arm-level physical human-robot interaction (Rouxel et al., 2023, Rouxel et al., 2024, Totsila et al., 2024, Razmjoo et al., 2023, Saqib et al., 30 May 2026).
The role of TALOS in these papers is not merely that of a generic hardware demonstrator. It is used as a physically demanding validation platform for methods that couple high-level decision making with whole-body feasibility, contact-force regulation, and real-time control. In upper-body multi-support manipulation, for example, only 22 joints are actively controlled, the feet remain fixed, and the right hand can be used as an additional environmental support contact; in the SEIKO whole-body-control experiments, a modified configuration uses optimized joints after removing the right forearm joints and ignoring head joints (Rouxel et al., 2024, Rouxel et al., 2023). In assistive sit-to-stand experiments, the control design is restricted further to 14 active joints with left-right symmetry assumptions (Razmjoo et al., 2023).
A plausible implication is that TALOS functions as a common embodiment for studying the interface between geometric planning, contact reasoning, and physical execution under strong hardware constraints. The recurring emphasis on fixed feet, support-hand contacts, explicit balance margins, and interaction safety suggests that the platform is frequently used near the boundary where posture, contact, and stability must be treated jointly rather than as separate problems.
2. Whole-body control, contact regulation, and motion generation
A central line of TALOS research concerns multi-contact whole-body control on a position-controlled humanoid. The SEIKO pipeline—Sequential Equilibrium Inverse Kinematic Optimization—formulates whole-body retargeting from Cartesian commands and admittance control using two quadratic programs solved in real time. Its quasi-static equilibrium model is written as
and its flexibility model uses
so that contact-force regulation becomes indirect but explicit even on a position-controlled robot. The full pipeline runs at 500 Hz, with commanded positions interpolated at 2 kHz, and is validated on real TALOS in pushing tasks, far-reaching tasks, stair climbing, and stepping on sloped surfaces (Rouxel et al., 2023).
A complementary direction combines learned high-level strategy with model-based whole-body feasibility. In “Flow Matching Imitation Learning for Multi-Support Manipulation” (Rouxel et al., 2024), a hierarchical system couples a Flow-Matching policy to the SEIKO retargeting and control stack. The policy maps a structured state to a future trajectory of Cartesian effector poses and continuous contact-state values over a horizon of steps, corresponding to $6.4$ s, with outputs generated at 5 Hz, stitched online, resampled to 100 Hz, and executed by the low-level stack at 500 Hz. On TALOS, this architecture is used for real dishwasher-drawer closing and non-prehensile box pushing, with the right hand acting as support and the left hand as the main manipulating effector (Rouxel et al., 2024).
The same pattern appears in locomotion-oriented planning. “NAS: N-step computation of All Solutions to the footstep planning problem” reports an algorithm that computes all possible solutions to a contact-planning problem, claims a globally optimal policy efficiently queried in real time, and states that NAS is demonstrated in a variety of scenarios for the Talos robot, both in simulation and on the hardware platform (Wang et al., 2024). Even though the detailed text provided for that paper is incomplete, its abstract places TALOS explicitly within the evaluation of globally complete contact-planning methods.
Taken together, these works position TALOS as a platform where contact-state generation, whole-body retargeting, and force-aware execution are co-designed. This suggests that the platform’s research significance lies less in any single controller than in enabling integrated studies of command generation, feasibility enforcement, and contact establishment on real hardware.
3. Language guidance, physical assistance, and safe interaction
Another TALOS-centered strand studies how humans specify or influence robot behavior. “Words2Contact: Identifying Support Contacts from Verbal Instructions Using Foundation Models” presents a language interface for selecting support contacts before whole-body reaching or manipulation. The pipeline distinguishes Prediction, Correction, and Confirmation requests; uses 5-shot prompting; maps an initial verbal instruction and RGB image to a 2D image-space contact point; supports iterative correction; and then converts the confirmed 2D point into a 3D target from the point cloud, with transformation
Execution is handled by SEIKO Retargeting and the SEIKO whole-body controller, and the real-world validation is explicitly performed on TALOS in tasks such as placing a hand on the top of a book, on a white surface, near a mallet on a table, or on white cloth over a fridge in order to avoid falling while reaching for distant objects (Totsila et al., 2024).
Physical assistance is addressed from a different angle in “Learning Joint Space Reference Manifold for Reliable Physical Assistance” (Razmjoo et al., 2023). That study uses TALOS for assistive sit-to-stand and stand-to-sit interaction, where a human can exert 100–200 N over 2–8 s. The proposed control reference is a 1D manifold in the space of 14 active joints, parameterized with Bernstein polynomials using 11 control points, and indexed online by the normalized horizontal interaction force
The manifold is optimized offline with a robust ZMP-boundary cost over a force range up to $99.7$0, then tracked online with a simple force-to-posture mapping. The experiments keep the feet fixed, constrain the hands to move no more than 10 cm from the default hand position, and use a conservative sagittal ZMP margin of $99.7$1 (Razmjoo et al., 2023).
Safe pHRI is also studied at the arm level. “Adaptive PD Gains for Energy-Conscious Control in Physical Human-Robot Interaction” implements an energy-limiting adaptive PD controller on the 7 joints of the right arm of a torque-controlled TALOS. The controller adapts gains based on total mechanical energy
$99.7$2
with separate handling of kinetic energy to increase damping under high-velocity disturbances. In hardware, the paper reports a nominal energy limit of $99.7$3 J; a traditional PD controller reaches $99.7$4 J, whereas the adaptive controller stays below the limit, increases joint-1 deflection from $99.7$5 rad to $99.7$6 rad, and increases settling time from $99.7$7 s to $99.7$8 s after release (Saqib et al., 30 May 2026).
These studies present TALOS as a platform for supervisory human input at multiple abstraction levels: natural-language contact specification, measured-force-driven posture adaptation, and safe compliant response to physical disturbance. A plausible implication is that TALOS is used not only to test autonomous control policies, but also to study how high-level human intent can be translated into physically safe whole-body behavior.
4. Embodiment assumptions and comparative mechanical characterization
The TALOS literature is not uniform about low-level actuation assumptions. In the SEIKO and support-contact works, TALOS is treated as a position-controlled humanoid robot whose joint flexibility and non-ideal position control can be exploited for indirect contact-force regulation (Rouxel et al., 2023, Totsila et al., 2024). By contrast, the adaptive pHRI study describes its implementation explicitly on a torque-controlled TALOS humanoid robot through ros_control, focusing on the right arm rather than on full-body multi-contact behavior (Saqib et al., 30 May 2026). This does not by itself establish a contradiction; rather, it indicates that different studies expose different control interfaces or subsystems of the platform.
TALOS also appears as a reference design in lower-body comparative mechanics. In “Modeling and Numerical Analysis of Kangaroo Lower Body based on Constrained Dynamics of Hybrid Serial-Parallel Floating-Base Systems”, TALOS is described as another full-size humanoid bipedal system, the TALOS robot, developed by PAL Robotics, with 6 actuated DOFs per leg, a more traditional serial kinematics design for its legs, and motors distributed along the entire chain. In that comparison, TALOS’s center of mass is described as being at knee height, whereas Kangaroo’s is closer to the pelvis. The paper compares the robots using equivalent Cartesian inertia at the foot and Centroidal Angular Momentum Matrix metrics, reporting TALOS/Kangaroo ratios of 3.6, 4.1, 4.6 for the linear part of equivalent Cartesian inertia, 11.4, 4.9, 3.6 for the angular part, and 2.6, 2.7, 2.0 for CAMM column norms (Hoffman et al., 2023).
That same paper adds a useful architectural nuance: in its broader humanoid-family table, TALOS (2017) is listed with 2 parallel mechanisms of 1-DOF, 27 total free DOFs, and 2 parallel free DOFs (Hoffman et al., 2023). This suggests that, although TALOS is used as a serial-leg reference in the lower-body comparison, its overall morphology is not purely reducible to a single simple kinematic label.
5. Acronymic uses of TALOS beyond the humanoid robot
Outside humanoid robotics, TALOS and its capitalization variants denote a large set of unrelated research artifacts. The acronym expansions are domain-specific and should not be conflated.
| Name | Expansion or role | Domain |
|---|---|---|
| TALoS | Test-time Adaptation via Line of Sight | Semantic Scene Completion (Jang et al., 2024) |
| Talos | privacy-preserving contrastive learning mechanism | Machine learning privacy (He et al., 2021) |
| Talos | defense based on global graph homophily | GNN robustness (Li et al., 2024) |
| Talos | efficient rack-scale LSM-based KV store | Storage systems (Vardoulakis et al., 2021) |
| TALOS | Toolbox for Analysis and Large-scale Optimization of Spacecraft | Spacecraft MDO (Gandarillas et al., 2023) |
| TALOS | Total Automation of LabVIEW Operations for Science | Experimental automation (Volponi et al., 2024) |
| TALOS | lightweight framework for verifiable state management and trustworthy application migration | TEE live migration (Vasileaidis et al., 5 Sep 2025) |
| Talos | system for Security Workarounds for Rapid Response | Software security (Huang et al., 2017) |
| TaLoS | Task-Localized Sparse Fine-tuning | Model editing (Iurada et al., 3 Apr 2025) |
| Talos | loss for Top-$99.7$9 recommendation accuracy | Recommender systems (Zhang et al., 27 Jan 2026) |
The substantive meanings of these names differ sharply. In semantic scene completion, TALoS uses line-of-sight supervision from temporally aligned LiDAR frames to adapt a pretrained SSC model at test time (Jang et al., 2024). In contrastive learning, Talos is an adversarially trained privacy-preserving mechanism that reduces attribute inference leakage from representations while preserving utility and membership privacy (He et al., 2021). In graph learning, Talos removes suspicious edges by maximizing a notion of global graph homophily rather than relying only on local edge similarity (Li et al., 2024). In storage, Talos is an RDMA-enabled replication design for LSM key-value stores that avoids compactions on backups by sending pre-built indexes from the primary (Vardoulakis et al., 2021).
The same acronym also names large software frameworks. TALOS in spacecraft design is the Toolbox for Analysis and Large-scale Optimization of Spacecraft, built on CSDL for large-scale multidisciplinary design optimization (Gandarillas et al., 2023). TALOS in experimental control is Total Automation of LabVIEW Operations for Science, an actor-based framework developed in the AEgIS collaboration for autonomous operation of complex experiments (Volponi et al., 2024). In confidential computing, TALOS denotes a framework for secure live migration of trusted applications in TEEs with verifiable state management and Proof-of-Execution-style continuity checks (Vasileaidis et al., 5 Sep 2025). The cybersecurity paper “Talos: Neutralizing Vulnerabilities with Security Workarounds for Rapid Response” uses the name for a source-code instrumentation system that inserts Security Workarounds for Rapid Response into server programs (Huang et al., 2017). More recently, TaLoS names a sparse fine-tuning method for composable task vectors in model editing (Iurada et al., 3 Apr 2025), and Talos also names a Top-$32$0-aware recommendation loss based on quantile thresholding and a tailored surrogate (Zhang et al., 27 Jan 2026).
6. Significance and disambiguation
The research record therefore treats TALOS less as a single concept than as a polysemous technical name. In robotics, it most often denotes the PAL Robotics humanoid and serves as a real-hardware substrate for studying whole-body retargeting, multi-contact force regulation, language-grounded support selection, learned contact switching, assistive posture manifolds, and compliant energy-limiting control (Rouxel et al., 2023, Rouxel et al., 2024, Totsila et al., 2024, Razmjoo et al., 2023, Saqib et al., 30 May 2026). In other fields, it denotes independent algorithms, toolboxes, or frameworks whose only commonality is the name.
A common source of confusion is capitalization. TALOS, Talos, TALoS, and TaLoS identify unrelated artifacts, often in completely different domains. The literature therefore distinguishes between the TALOS humanoid robot and acronymic constructs such as TALoS for line-of-sight test-time adaptation, TaLoS for task-localized sparse fine-tuning, or TALOS for spacecraft or laboratory-control frameworks (Jang et al., 2024, Iurada et al., 3 Apr 2025, Gandarillas et al., 2023, Volponi et al., 2024).
Within robotics specifically, TALOS is significant because it appears repeatedly at the boundary between high-level autonomy and low-level physical feasibility. The platform is used for experiments where contact creation, balance, force distribution, and interaction safety cannot be abstracted away. Across the cited work, TALOS is thus less a single method than a recurring embodiment through which contact-rich humanoid control, perception, and interaction are made experimentally concrete.
