HY-Embodied-0.5: Hybrid Embodied Intelligence

• HY-Embodied-0.5 is a framework for integrating multimodal perception, reasoning, and actuation through hybrid and situationally-aware architectures.
• It builds on embodied cognition and somatic psychology principles to enable both human-in-the-loop art therapy and advanced robotic control.
• The framework leverages specialized model architectures and training paradigms to achieve state-of-the-art benchmarks in spatial and embodied intelligence.

HY-Embodied-0.5 is a term used across embodied intelligence literature for distinct approaches that integrate hybrid perception, reasoning, and actuation through multimodal, situationally aware architectures. It characterizes both a specific family of foundation models for embodied agents and a hybrid embodied framework at the intersection of tactile, physiological, and digital modalities. Across these research lines, HY-Embodied-0.5 denotes systems engineered to bridge high-level perception—often spanning vision, language, and bodily states—with task-relevant outputs for both artificial and human-agent scenarios (Nasri et al., 15 Dec 2025, X et al., 8 Apr 2026).

1. Theoretical Foundations and Scope

HY-Embodied-0.5 commonly rests on the conceptual pillars of embodied cognition and somatic psychology in human-in-the-loop settings, as well as the need for high-fidelity multimodal integration, flexible policy learning, and spatial reasoning in real-world machine agents.

• In the context of mixed-reality art therapy, it leverages theories from Varela’s Embodied Mind and Craig’s interoceptive neuroscience, operationalizing emotional states as quantifiable physiological signatures (e.g., breath, HRV, ocular dynamics) and translating these signals into visual, spatial, and material artifacts for reflection and self-regulation (Nasri et al., 15 Dec 2025).
• For embodied foundation models, the focus is on equipping agents with the architectural prerequisites for spatial and temporal visual perception, advanced reasoning, and adaptive planning, via tightly integrated multimodal transformer networks and iterative training strategies (X et al., 8 Apr 2026).

This dual orientation—simultaneous grounding in bodily, affective processes and state-of-the-art computational modeling—enables HY-Embodied-0.5 systems to serve as bridges between subjective somatic experience, analog/digital synthesis, and robust action policy formation.

2. System and Model Architectures

Hybrid Embodied Art Framework

The HY-Embodied-0.5 hybrid framework features parallel analog and mixed-reality modalities:

• Analog Track: Clay sculpting and free-form drawing. Human interoception (body-scan, tension, breath) is rendered as spatial and material properties in the artifact, with direct physical feedback and archival.
• Mixed Reality Pipeline: Sensor acquisition (respiratory belt, PPG, 6-DoF tracker, eye-tracker), real-time signal processing, parameter mapping, and avatar rendering are orchestrated in environments such as Unity3D. Visual and spatial cues are actuated through parameterized functions: for example,

$$S_\mathrm{color}(t) = \alpha \times \frac{r(t) - r_{\min}}{r_{\max} - r_{\min}}$$

with similar constructs for pulsing, motion amplitude, and mesh deformation, creating a closed loop between biosignals and immersive feedback (Nasri et al., 15 Dec 2025).

Embodied Foundation Models

For machine agents, HY-Embodied-0.5 specifies two primary variants:

Model	Activated Params	Type	Use Case
MoT-2B	2B (~4B total)	Mixture-of-Transformers	Edge/real-time
MoE-A32B	32B (~407B total)	Mixture-of-Experts/MoT	Server/high-complexity

Both architectures use a native-resolution vision transformer (HY-ViT 2.0 for MoT-2B) with LLM backbone, employing modality-specific QKV/FFN block duplication for vision/text separation and multimodal attention masks. Latent tokens serve as visual-language bottlenecks, explicitly aligned to ViT teacher CLS features for improved perceptual representation.

The pre-training objective includes: $$\mathcal{L}_\mathrm{total} = \mathcal{L}_\mathrm{LLM} + \mathcal{L}_\mathrm{vision} + \mathcal{L}_\mathrm{global}$$ where each component addresses language, vision, and alignment constraints respectively (X et al., 8 Apr 2026).

3. Training Paradigms and Adaptation Strategies

HY-Embodied-0.5 employs multi-stage, self-evolving post-training paradigms:

• Supervised Fine-tuning (SFT): Chain-of-thought (CoT) reasoning sequences are generated and filtered for quality, bootstrapping multi-step, high-complexity reasoning pathways.
• Reinforcement Learning (GRPO): Task-specific reward functions, including geometric, trajectory, and regression-based signals, are standardized and normalized by group-relative advantage.
• Rejection-Sampling Finetuning (RFT): Model rollouts are frontier-sampled and scored by a stronger teacher, iteratively consolidating advanced CoT behaviors.

After refinement, on-policy distillation (OPD) is used to match the compact model’s token-level policy distributions to those of the large expert model on its own sampled trajectories: $$\mathcal{L}_\mathrm{OPD} = \mathbb{E}_{x, y \sim \pi_s}\left[\frac{1}{|y|}\sum_{t=1}^{|y|} \mathrm{KL}(\pi_t(\cdot\mid x, y_{<t}) \| \pi_s(\cdot\mid x, y_{<t}))\right]$$ This procedure ensures the student model inherits and generalizes the embodied reasoning competencies of the large teacher (X et al., 8 Apr 2026).

4. Evaluation and Benchmarking

HY-Embodied-0.5 models are evaluated across extensive benchmarks optimized for embodied intelligence:

• Visual Perception (e.g., CV-Bench, DA-2K): Classification and depth accuracy tasks.
• Spatial Understanding (e.g., 3DSRBench, MindCube): Fine-grained reasoning for geometric, multi-view, reference, and spatial constraint questions.
• Embodied Understanding: Downstream reasoning and action—robotics affordance, multi-choice planning, trajectory alignment.

Quantitative summary of benchmark results (micro-averages):

Model Variant	Average Score (%)	SOTA Comparison	No. of Wins (22 tasks)
MoT-2B	58.0	Qwen3-VL-4B (47.8)	16/22 (best), 4/22 (2nd)
MoE-A32B	67.0	Gemini 3.0 Pro (63.6)	7/22 (best), 6/22 (2nd)

In embodied robot control (Vision-Language-Action, VLA) transfer settings, MoT-2B outperforms prior policies (π0, π0.5) on challenging real-world manipulation tasks, especially under high-complexity object arrangement (X et al., 8 Apr 2026).

5. Interpretation, Reflection, and Design Guidelines

Comparative Reflections

• HY-Embodied-0.5 implementations in human-in-the-loop art therapy reveal stronger embodied connection and self-reflective expressivity when feedback channels are slow and minimalist (≥200 ms latency, ≤3 channels), while rapid involuntary mirroring can overwhelm interoceptive focus (Nasri et al., 15 Dec 2025).
• In machine scenarios, Mixture-of-Transformers architectures with explicit multimodal channel separation and latent bridging (vision→language) consistently outperform similarly sized VLMs, with large models matching or beating frontier baselines on spatial/embodied benchmarks (X et al., 8 Apr 2026).

Design Constraints and Workflow Principles

Human-facing hybrid frameworks:

• Maintain tunable mappings to preserve user agency.
• Limit concurrent visual/spatial feedback to mitigate cognitive overload.
• Anonymize stored biometric data using only visual/spatial abstraction.
• Initiate all workflows with body-scan meditation and allow analog fallback modes for trauma-sensitive users (Nasri et al., 15 Dec 2025).

Embodied model architectures:

• Employ modality-specific QKV/FFN duplication within transformers for explicit channel control.
• Use tight bottlenecks (latent tokens) with explicit alignment for vision–language compositional transfer.
• Refine and consolidate reasoning trajectories via alternating SFT/RL/RFT cycles before distillation to compact agents (X et al., 8 Apr 2026).

6. Representative Applications and Extensions

• Hybrid somatic/MR pipelines for emotional archiving and therapeutic workflows, supporting trauma-informed journaling with revisitable 3D artifacts.
• Edge and server-scale generalist agents for spatial, visual, and temporal reasoning, with robust transfer for downstream robot control.
• Embodied scene description agents, using hybrid IL+RL, hierarchical skills, and perception upgrades, as demonstrated in real room navigation and semantic description transfer (Tan et al., 2020).
• Adaptive policy managers for swarm robot embodiment and embodied flight, leveraging fine-grained affordance and intuitive feedback control (Ichihashi et al., 2024, Cherpillod et al., 2017).

7. Future Directions and Open Questions

While HY-Embodied-0.5 demonstrates state-of-the-art performance and cross-domain applicability at maturity level 0.5, several open research avenues remain. These include the exploration of deeper hierarchical planning, further reduction of perception–action latency, ethical archiving of biometric data in human settings, and the expansion of the design vocabulary for mapping internal and external bodily states to expressive digital or robotic affordances (Nasri et al., 15 Dec 2025, X et al., 8 Apr 2026). A plausible implication is that continued refinement in hybrid analog-digital workflows and scaling of embodied model architectures will catalyze new directions in both human and artificial embodied cognition.