X Robotic Model 1 (XR-1)

• XR-1 is a dual-concept robotic platform combining a wearable augmentation device for hazardous payload support with a vision-language-action model for multi-robot manipulation.
• The wearable system enhances mobility by redistributing loads and optimizing joint torques, enabling dynamic transitions such as standing, crouching, and crawling.
• The VLA model employs dual-branched VQ-VAE to align visual and motion cues, achieving robust generalization across diverse robotic embodiments.

X Robotic Model 1 (XR-1) refers to two distinct research-grade robotic systems published under the same system moniker: (1) a wearable lower-limb augmentation device for human payload support in hazardous environments, and (2) a scalable vision-language-action (VLA) model for generalist policy learning across heterogeneous robot embodiments. Both are rigorously documented in the academic literature, yet target fundamentally different aspects of robotics—physical augmentation and unified AI-driven manipulation.

1. System Objectives and Scope

XR-1 as Human Augmentation Platform:

Originally introduced as the “Extra Robotic Legs” (XRL) system, XR-1 was conceived as a wearable, articulated two-legged robotic platform aimed at liberating emergency response personnel from the load and postural fatigue imposed by heavy personal protective equipment (PPE) (Gonzalez et al., 2020). Its functional objectives include:

• Maintaining upright balance while supporting a 22.7 kg payload on the operator’s back.
• Delivering a 222.4 N upward assistive force to the torso during crouch and crawl, in conjunction with payload carriage.
• Enabling seamless transitions between standing, crouching, and crawling without interrupting ongoing tasks.
• Supporting walking, stair climbing, and postural transitions with full external load transfer through the exoskeleton.
• Sustaining continuous operation for at least one hour in emergency-response contexts.

These objectives dictate scenario coverage spanning upright walking, deep crouching, ground-level crawling, and stair negotiation—removing the biomechanical burden of SCBA tanks, tool belts, and additional PPE through full robotic kinematic transfer of loads.

XR-1 as Vision-Language-Action Framework:

A subsequent usage of XR-1 designates a unified VLA model designed to bridge visual and linguistic instructions to low-level motor actions across diverse robots and environments (Fan et al., 4 Nov 2025). This version is constructed to:

• Map high-dimensional sensory signals (multi-view images, proprioception, language) onto precise joint and actuator commands.
• Generalize across robot morphologies and manipulation tasks using embodiment-agnostic latent codes.
• Leverage both large-scale robotic demonstrations and human activity video for scalable representation learning.
• Deliver robust performance in real-world manipulation settings, including unseen objects, dynamic distractors, and varied scene backgrounds.

2. Mechanical and Algorithmic Architecture

Mechanical Structure and Kinematics (Augmentation Platform):

Each XR-1 leg implements a 6-DOF serial chain, with a differential-powered hip, single-DOF knee, and differential-powered ankle (pitch/roll). The system attaches rigidly at the waist via a 6-axis force-torque interface. To minimize actuator loads in demanding postures, segment lengths are set so full-upright and full-crouch configurations lie near kinematic singularities, locking under gravity and minimizing continuous motor torque.

Formally, let $\ell_1$ and $\ell_2$ be the primary segment lengths, with design constraints:

$$\ell_s = \ell_1 + \ell_2, \qquad \ell_c = \ell_2 - \ell_1$$

where $\ell_s$ and $\ell_c$ correspond to hip heights in standing and crawling, respectively.

In transitions (notably the squat), a closed-loop frontal-plane posture permits internal force and moment redistribution:

• Joint torques for a planar 3-link arm are given by $\tau = J(\theta)^\top W$.
• In frontal squat, internal parameters $[F_y, M]$ minimize peak actuator torque under a minimax objective, yielding an optimal redistributed solution that numerically reduces maximum per-joint loading from $>300$ Nm to $\approx 30$ Nm in squat postures.

Unified Vision-Motion Representation (VLA Model):

The core algorithmic innovation is the “Unified Vision-Motion Codes” (UVMC, Editor's term), obtained via a dual-branch VQ-VAE:

• The visual encoder $E_{\rm vis}$ maps frame pairs $(\bcamera_t, \bcamera_{t+h})$ to latent $z_{\rm vis}$; its decoder reconstructs future frames.
• The motion encoder $E_{\rm mo}$ maps action-proprioceptive tuples $(\ba_{t:t+h}, \bmea_{t:t+h})$ to $z_{\rm mo}$; its decoder reconstructs action sequences.
• Both project through a shared codebook $\mathcal{C}$ of discrete embeddings, enforced via VQ-VAE quantization and a KL-based cross-modal alignment regularizer:

$$\mathcal{L}_{\rm align} = D_{\rm KL}( q(z_{\rm mo}^e) \,\|\, q(z_{\rm vis}^e) )$$

The concatenated quantized codes $(z_{\rm vis}^e, z_{\rm mo}^e)$ form UVMC tokens mediating between multi-modal perception and action.

3. Actuation, Transmission, and Control Methods

Physical Actuation (Augmentation Platform):

To support high-torque yet back-drivable behavior, XR-1 uses brushless outrunner motors (15-pole, 0.45 Nm/A, up to 22.5 Nm @ 50A), organized in differential pairs at major joints. Transmission employs spiral miter gears—motor pairs drive a single output shaft, doubling effective torque and supporting torque preloading. Gear reduction remains below 10:1, preserving high bandwidth force/torque control.

A summary of actuation specifications is presented below:

Parameter	Value	Description
Motor torque constant ($k_\tau$)	0.45 Nm/A
Peak torque per motor	22.5 Nm	2 × per joint, via differential
Gear reduction	≈ 1:2 (per motor)
Joint torque (2 motors)	Up to 45 Nm	Without further gearing

Torque control loops are implemented onboard at 10 kHz, leveraging Hall-effect and magnetic encoder feedback. The “virtual ankle” control scheme stabilizes the overall system by rapid torque modulation in response to deviations in the center of mass.

Learning and Inference (VLA Model):

XR-1 (VLA) operates as follows:

1. At each timestep, observations $(\bcamera_t, \bmea_t)$ and instruction $\blang$ are processed by a vision-LLM $F(\blang, \bobs_t)$, extended with two learnable UVMC tokens.
2. The action prediction head $H(\cdot)$, conditioned on these multimodal features and UVMC, outputs the next robot action $\hat{\ba}_t$.
3. During training, $F$ is supervised to predict UVMC codes and reconstruct actions, with domain-aligned pretraining spanning multiple robot types.
4. At deployment, only $F$ and $H$ are active; UVMC tokens encode the multimodal priors necessary for broad generalization.

4. Training Paradigms and Evaluation Protocols

Three-Stage VLA Training Curriculum:

• Stage 1: Self-supervised UVMC pretraining on robot demonstrations and human videos to learn the discrete latent representation.
• Stage 2: UVMC-guided pretraining in a large-scale VLA model, aligning features across robots and tasks.
• Stage 3: Task-specific post-training for embodiment specialization.

In total, >1.2M episodes and 110M frames were used for UVMC pretraining, spanning datasets such as Open-X, RoboMIND, XR-D, and Ego4D.

Empirical Metrics:

• Evaluations on six robot embodiments (including Tien Kung, UR-5e, Franka, AgileX Magic) and 120 manipulation tasks.
• XR-1 demonstrates average success rates such as 75.3% (Single-Arm UR-5e), 73.5% (Dual-Arm Franka), versus state-of-the-art baselines ($\pi_{0.5}, \pi_0$, RDT, UniVLA, GR00T-N1.5).
• Out-of-box generalization (without task-specific tuning) on 7-task Dual-Arm Franka yields parity or better performance relative to baselines, with success rates under background shifts, distractions, and novel objects remaining within 55–65% (vs. 5–50% for non-UVMC models).

Prototype Validation (Augmentation Platform):

• Joint-level step torque commands (0→30 Nm) achieve <5 ms rise time.
• Five-minute squat with 50 lb assistive force induces <2 °C thermal rise at the motors.
• Closed-loop torque redistribution in squat posture shows ~85% reduction in knee motor current compared to sagittal configuration.

5. Insights, Limitations, and Future Directions

Ablation studies on the VLA XR-1 indicate:

• UVMC fine-tuning alone (exempting Stage 1) lifts average task success from 28.3% to 48.3%.
• Applying KL-based vision-motion alignment further increases accuracy to 66.7%.
• Scaling UVMC pretraining enhances monotonic gains; injection of cross-embodiment data in Stage 2 delivers success rates up to 81.6%.
• Lightweight model variants (“XR-1-Light” at 230M params) preserve UVMC benefits under resource constraints.

Physical lessons from the augmentation platform include requirements for improved hip yaw actuation, harness ergonomics, and integrated onboard power for hour-level runtime. Planned XR-2 improvements will address dynamic gait, human intention estimation, and full tetherless operation.

A plausible implication is that UVMC provides an effective, compact bottleneck to align high-dimensional perception with actionable motor primitives, enabling broad generalization across tasks and embodiments. Similarly, exploiting closed-loop kinematics and redundancy in physical exoskeletons permits significant mechanical efficiency gains.

6. Comparative Analysis and Implications

The dual conception of XR-1 as both augmentation hardware and as a unified VLA model reflects broader trends in robotics: hardware-software co-design, embodiment-agnostic representation learning, and scalability to complex, real-world scenarios.

When compared to prior state-of-the-art robotic control systems:

• The augmentation-focused XR-1 demonstrates a reduction in required actuator torque by an order of magnitude via posture and kinematic strategy, crucial for power and weight constraints in wearables.
• The VLA-based XR-1 outperforms contemporary multi-robot learning architectures in both systematic manipulation tasks and OOD settings, confirming the necessity of joint visual-motion codebooks and large-scale cross-modal pretraining.

Open challenges remain in dynamic whole-body control, tetherless power delivery for wearable augmentation, and scaling VLA models to continually evolving tasks. Cross-modal alignment in the UVMC approach appears essential for unlocking true robot generalist capabilities in real settings.