Husky Beta in Robotics and Language Agents

• Husky Beta is a context-dependent designation used to denote distinct systems: a quadrupedal robot prototype for advanced locomotion and an open-source language agent for multi-step reasoning.
• In robotics, the Husky Beta variant employs structure repurposing with dual-mode (legged-aerial) locomotion and thrust-assisted stabilization, highlighting precise posture manipulation and GRF coordination.
• The language agent version utilizes a modular action decomposition with integrated expert models, iterative inference loops, and extensive tool-based reasoning to achieve robust performance.

Searching arXiv for the cited Husky-related papers to ground the article in current preprints. Husky Beta is a context-dependent designation rather than a single canonical system in the arXiv literature. In Northeastern University robotics papers, it denotes the carbon-based, off-the-shelf–actuated quadrupedal legged–aerial prototype also called Husky v.2, developed for structure repurposing, thruster-assisted narrow-path locomotion, and steep-slope walking (Wang et al., 10 Oct 2025). In language-agent research, “Husky Beta” is used as an interpretive label for the public v1 release of Husky, a unified open-source language agent for multi-step reasoning (Kim et al., 2024). By contrast, the humanoid skateboarding system HUSKY does not describe any explicit “Husky Beta” version or staged release (Han et al., 3 Feb 2026). This suggests that the term functions primarily as a local name whose meaning must be inferred from research domain and paper lineage.

1. Terminological scope and naming

The term appears in at least three distinct usages in the supplied literature.

Usage context	Referent	Status of “Husky Beta”
Northeastern University multimodal robotics	Husky v.2 quadrupedal legged–aerial robot	Internal moniker for the v.2 prototype (Wang et al., 10 Oct 2025)
Open-source language agents	Public v1 release of Husky	Interpreted as “beta,” but not explicitly labeled as such (Kim et al., 2024)
Humanoid skateboarding	HUSKY: humanoid skateboarding framework	No explicit “Husky Beta” release described (Han et al., 3 Feb 2026)

In the Northeastern lineage, the naming is comparatively direct: Husky Beta = Husky v.2, and the platform is described as the successor to Husky Carbon v.1 (Wang et al., 10 Oct 2025). In the language-agent literature, the paper does not explicitly tag the release as beta, but it provides code and trained models and presents clear training and inference procedures, ablations, and extensive evaluation, which motivates the “beta” interpretation (Kim et al., 2024). In the skateboarding paper, the opposite is true: the paper and project page do not mention or define a “Husky Beta” release, package, or staged deployment (Han et al., 3 Feb 2026).

A common misconception is therefore to treat “Husky Beta” as a unique research artifact. The literature instead supports a disambiguated reading in which the name indexes unrelated systems: a legged–aerial quadruped, a tool-augmented language agent, and, only orthographically, a separate humanoid skateboarding framework.

2. Husky Beta as a unified language agent

In the language-agent setting, Husky is a modular, open-source language agent that decomposes a complex task into a sequence of discrete actions drawn from a unified action space and iterates between action generation and action execution via expert models (Kim et al., 2024). The unified tool set is $$\mathcal{T}=\{\texttt{[code]},\texttt{[math]},\texttt{[search]},\texttt{[commonsense]},\texttt{[finish]}\}$$, and the agent state is written as $s_t=(x,h_t)$, where $x$ is the task instruction and $h_t$ is the solution history. The action generator predicts $a_t=([\tau_t],\mathrm{step}_t)$ according to $a_t\sim\pi_\theta(a\mid s_t)$, and the state update appends the step and its resulting observation to the history.

The execution model is explicitly heterogeneous. For [code], a code generator writes Python, the program is executed, and a base reasoner rewrites the execution result into natural language. For [search], a query generator produces a search query, a Google search via SERP API returns an answer-box snippet or top organic result, and the result is rewritten to fit the task context. [math] produces direct natural-language mathematical reasoning, [commonsense] handles residual reasoning or tool-output rewriting, and [finish] emits the final answer. The inference loop is bounded by a strict maximum of 10 iterations, uses vLLM for efficient parallel execution, and sets temperature 0 for finetuned modules and 0.3 for the base reasoner (Kim et al., 2024).

The training pipeline is based on synthetic tool-integrated trajectories generated by a teacher LM. The action generator is trained on 110K instances extracted from tool-integrated trajectories across numerical, tabular, knowledge, and mixed domains; the code generator uses 44K code snippets; the math reasoner uses 30K math solution instances; and the query generator uses 22K search-query instances (Kim et al., 2024). Across the trajectory corpus, the paper reports 51,461 total trajectories, 35,137 correct trajectories retained, and 111,330 actions extracted. Training uses DeepSpeed ZeRO-3, BF16, 3 epochs, learning rate $5\times10^{-6}$, weight decay $10^{-2}$, warmup ratio 0.03, max length 2048, and total batch size 32.

The evaluation emphasizes breadth rather than task specialization. Husky is evaluated zero-shot on 14 datasets spanning numerical, tabular, knowledge, and mixed-tool reasoning. Reported scores include 79.9 accuracy on GSM-8K for the Llama-3 8B variant, 42.1 on MATH, 77.6 EM on TabMWP for Husky-7B, 58.4/70.2 EM/F1 on Bamboogle for Husky-8B, and 25.0 accuracy on HuskyQA for Husky-13B (Kim et al., 2024). On mixed-tool tasks, the paper reports that Husky-7B surpasses GPT-4 on DROP* and IIRC*, and that Husky-13B reaches 25.0 on HuskyQA. The accompanying benchmark HuskyQA is specifically constructed to require retrieval of missing facts and subsequent numerical reasoning.

The system’s limitations are likewise explicit. Retrieval can be brittle when a query under-specifies the entity, SERP API rate limits can constrain scale, the base reasoner is not fine-tuned and can propagate upstream errors, and the four-tool ontology may be insufficient for tasks requiring specialized APIs or richer document interaction (Kim et al., 2024). In this usage, “Husky Beta” therefore refers to a public, extensible, but still evolving open-source reasoning agent.

3. Husky Beta in the Northeastern quadruped lineage

In Northeastern University’s multimodal robotics literature, Husky Beta denotes the Husky v.2 prototype, a quadruped whose legs are repurposed as propeller arms for aerial mode (Wang et al., 10 Oct 2025). The platform is the successor to Husky Carbon v.1, which had custom-built actuators and earlier work on integrated thrust-vectoring and posture manipulation. Husky Beta reworks those ideas into a lighter system built with off-the-shelf servos and BLDC propulsors to emphasize controls and experimental validation of structure repurposing.

The hardware is defined by three actuated degrees of freedom per leg: hip frontal flexion/extension, hip sagittal flexion/extension, and knee flexion/extension. The knee drives a parallel four-bar linkage and mechanically carries a knee-mounted propeller. Hip frontal and hip sagittal joints use Dynamixel XH540-W270-T servos with stall torque 9.9 Nm, while knee joints use Dynamixel XM540-W270-T servos with stall torque 10.6 Nm (Wang et al., 10 Oct 2025). Propulsive elements are four SunnySky X4112S BLDC motors, each controlled by EOLO 50A Light ESCs. In the structure-repurposing paper the propellers are 14×4.7 double-blade props, whereas the narrow-path MPC paper reports 15×5.5 double-blade propellers for Husky β (Wang et al., 31 Jul 2025). The femur is an off-the-shelf carbon fiber tube, while tibia, fibula, ankles, foot linkages, and feet are 3D-printed in Markforged Onyx with fiber inlay.

The reported mass and component breakdown clarifies the design logic. The body mass is 1.68 kg; each leg contributes 0.698 kg, or about 2.8 kg across four legs; each BLDC motor is 0.197 kg; each propeller is 0.021 kg; each servo is 0.165 kg; and the battery mass is 0.50 kg total (Wang et al., 10 Oct 2025). The rationale for structure repurposing is precisely to avoid adding dedicated arms for flight, thereby reusing roughly 2.8 kg of leg structure as aerial appendages. On a 6s LiPo supply, maximum thrust is reported as 13.4 kg, with a resulting thruster-to-weight ratio approximately 2 for the test configuration (Wang et al., 10 Oct 2025).

The compute and sensing stack reflects the multimodal split. A Jetson Orin Nano runs high-level logic and leg controllers, a Cube Orange+ flight controller handles aerial loops, and communication uses pymavlink. Joint feedback comes from the servos, while the flight controller provides inertial sensing. Pinocchio is used for kinematic mapping from desired foot trajectories to joint commands (Wang et al., 10 Oct 2025). Earlier Husky Carbon configurations instead used Speedgoat Real-Time, ELMO Gold Twitter Solo amplifiers, RLS RMB20 encoders, VectorNav VN-200, Intel RealSense T265, and Pixhawk-based stabilization, indicating a hardware evolution from custom actuation and external control infrastructure toward a more self-contained off-the-shelf platform (Krishnamurthy, 2023).

A defining capability of the platform is posture manipulation and thrust vectoring hardware. Hip frontal rotation splays the legs outward so the femur–tibia linkage becomes approximately parallel to the floor; hip sagittal rotation then swings the knee-mounted propellers until their thrust axes become horizontal and the resultant center of lift aligns with the center of mass (Wang et al., 10 Oct 2025). In ground mode, the same thrusters can be used for roll stabilization during trotting or narrow-path walking. This makes “Husky Beta” in the robotics literature a platform concept as much as a single robot: a quadruped whose appendages are intentionally dual-use.

4. Reduced-order modeling and control formulations

Several papers develop control formulations around Husky-class platforms, and Husky Beta is central to the QP/MPC-based variants. For steep-slope walking, the platform called Husky Carbon; referred to here as Husky Beta is modeled with a modified Variable Length Inverted Pendulum (VLIP) under a fixed Zero Moment Point assumption, with a COM-collocated thruster and a QP MPC that solves for ground reaction forces and thruster forces (Krishnamurthy et al., 2024). In slope-aligned planar coordinates, the dynamics are

$$m\ddot{x}=-\lambda_x-mg\sin\alpha+F_x,\qquad m\ddot{y}=\lambda_y-mg\cos\alpha+F_y,$$

and the fixed-ZMP condition is expressed through the coupling between COM dynamics, COP, and GRF. The discrete prediction model uses horizon $n_h=5$ and time step $s_t=(x,h_t)$0, with qpSWIFT as the solver.

For narrow-path walking, an earlier Husky Carbon formulation uses the Husky Reduced-Order Model (HROM), in which the body is a single rigid body, the legs are treated as massless variable-length links, foot-ground contact uses a compliant normal force with Stribeck friction, and the thrusters are modeled as an external wrench on the body (Krishnamurthy et al., 2024). The reduced body dynamics are written as

$s_t=(x,h_t)$1

with leg actuation modeled as prescribed joint accelerations. The optimal control problem is then transcribed by cubic interpolation and midpoint collocation and solved with MATLAB fmincon, with the objective

$s_t=(x,h_t)$2

The later Husky β narrow-path paper replaces the HROM/collocation formulation with a centroidal-dynamics-based MPC/QP controller that regulates both GRFs and knee-mounted thruster forces (Wang et al., 31 Jul 2025). The centroidal model includes thruster directions that vary with leg posture:

$s_t=(x,h_t)$3

$s_t=(x,h_t)$4

The state is linearized and discretized as $s_t=(x,h_t)$5, the cost is quadratic in state-tracking and control effort, and the QP enforces dynamics consistency, linearized friction cones, stance/swing logic, and unilateral thruster bounds $s_t=(x,h_t)$6 (Wang et al., 31 Jul 2025).

Across these works, a stable modeling pattern emerges. Reduced-order representations are used to make thrust-assisted locomotion computationally tractable; contact forces and thrust are co-optimized or jointly scheduled; and friction constraints are central because the support geometry is either slope-limited or effectively line-like. A plausible implication is that Husky Beta serves as a laboratory for comparing several controller families—collocation-based optimal control, fixed-ZMP VLIP MPC, and centroidal QP/MPC—under the unifying theme of posture manipulation plus thrust assistance.

5. Demonstrated capabilities and reported performance

The structure-repurposing experiments report untethered two-contact trotting, push recovery, thruster-assisted roll stabilization in ground mode, and a full leg-to-flight transition followed by hovering for approximately 20 seconds (Wang et al., 10 Oct 2025). The documented sequence is: crouch, balance on a perch, splay the legs, rotate the knee-mounted props to horizontal, fix servo angles and power the propellers, then enter standard quadcopter hover. The transformation takes ~10 seconds, and the complete demonstration includes roughly 8 seconds of trotting, ~10 seconds of transition, and ~20 seconds of hovering.

The narrow-path MPC/QP study evaluates Husky β on a rigid beam of width 0.1 m and height 0.1 m (Wang et al., 31 Jul 2025). In beam walking, the reported thruster use is <7 N per thruster, substantially below the hardware maximum, while roll, pitch, and yaw remain within small bounds and ground friction ratios stay within no-slip limits. In the lateral push-recovery study, a 40 N lateral force is applied at the COM from $s_t=(x,h_t)$7 to $s_t=(x,h_t)$8, corresponding to a 20 N·s impulse. With thruster assistance and per-thruster bound 20 N, the robot recovers by $s_t=(x,h_t)$9; without thrusters, the gait fails by $x$0.

The steep-slope walking study reports feasibility on a $x$1 incline with friction coefficient $x$2, using the VLIP-QP MPC with $x$3 and $x$4 (Krishnamurthy et al., 2024). Solver performance is reported as ~65 microseconds per solve with a maximum of 10 iterations. The paper’s illustrative feasibility calculation states that for $x$5 and $x$6, traction without thrusters is insufficient, while a normal thruster component $x$7 suffices to raise the traction limit above the gravitational downslope component.

Earlier Husky Carbon narrow-path simulations establish the precursor envelope. The HROM/collocation paper reports that the body COM advanced about 0.3 m in 3.5 s, corresponding to an average speed of approximately 0.1 m/s, while body $x$8 and $x$9 positions remained stable and thruster commands remained within nominal EDF limits (Krishnamurthy et al., 2024). The thesis reports a 10-second thruster-assisted walking simulation that progressed approximately 1.6 m forward—about 0.16 m/s—with bounded roll error under thruster PID compensation, while hardware experiments executed open-loop walking with gait time 0.5 s and step length 0.1 m (Krishnamurthy, 2023).

Taken together, these results show that the Husky Beta lineage is not limited to one locomotor regime. The same design family has been studied for line-like support traversal, lateral push rejection, steep-slope ascent, and leg-to-hover transition. The precise controller differs across papers, but the experimental theme remains consistent: thrust is used not as a replacement for legs, but as an auxiliary source of stabilizing wrench when the support polygon, terrain friction, or gait dynamics make purely legged control insufficient.

6. Relation to the humanoid skateboarding system HUSKY

The humanoid skateboarding system “HUSKY: Humanoid Skateboarding System via Physics-Aware Whole-Body Control” is technically unrelated to the Northeastern Husky Beta quadruped and to the Husky language agent, despite the name overlap (Han et al., 3 Feb 2026). It addresses humanoid skateboarding on an underactuated, non-holonomic wheeled platform with hybrid contacts, models the coupling between board tilt and truck steering by

$h_t$0

and combines Adversarial Motion Priors, a heading-oriented physics-guided steering strategy, and trajectory-guided phase transitions.

The release status is explicit: the paper and project page do not mention or define a “Husky Beta” release, do not announce public code or pretrained policies at the time of writing, and specify only that reproduction details are sufficient to reimplement the system using mjlab + MuJoCo/IsaacLab, the simplified skateboard model, PPO, AMP training, physics-guided steering targets, and sim-to-real identification procedures (Han et al., 3 Feb 2026). The experiments are performed on the Unitree G1 humanoid rather than on any Husky-branded quadruped.

This distinction is important because the skateboarding paper can easily be conflated with the other Husky usages. The supplied evidence indicates the opposite: the skateboarding framework uses HUSKY as the project name, while Husky Beta is not a supported release designation in that paper. A plausible implication is that, within arXiv discourse, orthographic similarity alone is not enough to establish lineage or platform identity.

7. Research significance and open directions

Across domains, Husky Beta names systems that emphasize modularity under difficult control allocation. In the language-agent literature, modularity appears as a unified action space coordinated with specialized expert models; in the robotics literature, it appears as structure repurposing, posture manipulation, and thrust-assisted control under constrained contact conditions (Kim et al., 2024). The shared theme is not embodiment or application, but decomposition: each system factors a hard task into coordinated subsystems with explicit interfaces.

For the quadruped robotics lineage, the open problems are well specified. The structure-repurposing paper identifies compliance from lightweight Onyx/carbon construction, short hover endurance, perch-dependent transitions, and mode-separated control as current limitations, and proposes unified controllers integrating thruster inputs and GRFs, exploitation of posture manipulation in flight, payload addition, and further optimization of mass distribution and structural stiffness (Wang et al., 10 Oct 2025). The narrow-path MPC paper leaves onboard deployment, robust estimation under sensor noise, delay handling, and extension from rigid beams to flexible rope traversal as future work (Wang et al., 31 Jul 2025). The steep-slope study similarly highlights the promise of coordinated leg traction and COM-collocated thrust within a computationally efficient QP-MPC framework (Krishnamurthy et al., 2024).

For the language-agent usage, the forward path is also explicit: scale the action space and number of expert models, improve math and code specialists, and incorporate stronger general LMs while preserving the unified control loop (Kim et al., 2024). Because the code and models are already available, this branch of “Husky Beta” is the most immediately reproducible. By contrast, the skateboarding HUSKY branch presently offers reimplementation details but no described beta-stage release (Han et al., 3 Feb 2026).

The term therefore occupies an unusual place in contemporary research nomenclature. It does not identify a single benchmark, robot, or software package. Instead, it marks several technically serious but domain-separate programs of work: a public tool-augmented language agent, a family of multimodal legged–aerial robots centered on Husky v.2, and a frequently confused but distinct humanoid skateboarding framework. For arXiv readers, accurate interpretation of “Husky Beta” depends on tracing that lineage rather than on the name alone.