Lumo-2: Latent World-Action Model for Predictive, Aligned, and Scalable Robot Learning
Technical ReportToday, we introduce Lumo-2, our next-generation latent world-action model for generalist embodied robot learning.
Compared to Lumo-1, Lumo-2 delivers fundamental iterative upgrades across three core dimensions: it replaces explicit textual structured planning with implicit predictive reasoning grounded in latent world dynamics; it evolves from single-stage joint training to a three-stage progressive pre-alignment paradigm for cross-modal representation; and it substantially enhances scalability, shifting from reliance on dedicated robot demonstration data to native support for diverse multi-source data, with more favorable scaling properties and broader out-of-distribution generalization headroom.
These architectural upgrades collectively yield higher inference efficiency and stronger overall model performance. Lumo-2 achieves consistent improvements across challenging real-world tasks requiring temporal reasoning, physical understanding, and high-precision long-horizon and dexterous manipulation, making it better suited for practical deployment demands and custom development.
Core Technical Innovations
Built on three foundational principles — predictive reasoning, aligned representations, and scalable learning — Lumo-2 delivers a full architectural upgrade over Lumo-1, addressing core limitations of existing VLA and WAM systems.
Predictive reasoning in latent world dynamics
Replaces explicit textual planning with implicit predictive reasoning over a physically grounded latent space. This lightweight formulation encodes action-inducible future physical states, enabling world-modelling-style control without the overhead of dense pixel-level generation.
Multi-stage progressive modality pre-alignment
Resolves the persistent gap between action reconstruction fidelity and downstream control performance. Action representations are progressively aligned with world dynamics, vision and language, producing a semantically structured latent space with consistent cross-modal semantics.
Favorable scaling and generalization
The structured latent space yields improved scaling properties and stronger out-of-distribution generalization. The model natively supports efficient fine-tuning with diverse, accessible data sources including egocentric human videos.
Predictive reasoning in latent world dynamics
Moving beyond Lumo-1's explicit textual reasoning pipeline and the heavy computational overhead of pixel-level video world models, Lumo-2 constructs a lightweight, physically grounded latent dynamics space as its implicit reasoning substrate.
Rather than synthesizing dense future frames, the model encodes only action-relevant physical motion patterns in latent space, enabling world-modelling-style predictive reasoning with minimal compute cost. The learned representation exhibits native cross-embodiment semantic consistency: semantically equivalent behaviors across different robot platforms and human demonstrations converge to neighboring latent regions, forming a natural foundation for cross-platform generalization and human-to-robot transfer.
Multi-stage progressive modality pre-alignment
A defining advancement over conventional VLA/WAM designs, this paradigm resolves the longstanding mismatch between action reconstruction fidelity and real-world control performance. Following a curriculum of increasing alignment difficulty, action representations are progressively elevated from raw kinematic signals to semantically structured embeddings across three stages:
Stage 1: Aligning Actions with Latent World Dynamics
Establishes bimodal association between visual world dynamics and action representations via bidirectional constraints. Action representations are anchored to physical environment constraints, while the latent dynamics space is guided to focus on motion-relevant physical changes.
Stage 2: Aligning Actions with Vision and Language
Introduces semantic enhancement module and multi-task learning to further align action representations with the vision-language semantic space. This endows action tokens with linguistic interpretability and builds a direct semantic space mapping from natural language instructions to motor execution.
Stage 3: Co-Training on VLM, Video, and Robot Data
Performs multi-task joint training on top of the already aligned representation space, solidifying the causal reasoning structure of "predict future state first, generate actions second". This ensures all generated actions are grounded in physically plausible future outcomes.
Favorable scaling and generalization
Powered by the aligned latent space and predictive reasoning architecture, Lumo-2 delivers systematic improvements in inference efficiency, training scalability, and cross-domain transferability, meeting the core demands of industrial deployment.
Real-time inference performance
The block-wise autoregressive decoding strategy leverages the weak coupling across action dimensions to enable parallel generation within blocks, delivering a 2.71× end-to-end speedup over standard autoregressive generation with no accuracy loss, enabling closed-loop real-time control.
Robust long-horizon execution
A lightweight temporal context mechanism supplements sequential information via historical action memory, eliminating perceptual aliasing in multi-phase continuous tasks and markedly improving the continuity and success rate of long-horizon manipulation.
Broad data compatibility
The framework natively supports co-training on heterogeneous data sources including general VLM corpora, egocentric human videos, and multi-robot demonstration data, delivering predictable performance scaling with growing data scale.
Cross-domain transfer
No specialized transfer learning algorithms are required to achieve cross-embodiment and human-to-robot skill transfer, drastically reducing adaptation costs for new scenarios and robot platforms.
Experimental Evaluation
We conduct a comprehensive empirical evaluation to validate Lumo-2 along four core research dimensions.
Embodied VLM Benchmarks
Lumo-2 is evaluated on two families of embodied benchmarks: embodied understanding and embodied location, with unified evaluation configurations. Compared to the base Qwen-3.5 4B model, Lumo-2 VLM achieves consistent and significant improvements across nearly all tasks. It also delivers highly competitive performance against other state-of-the-art 4B-scale embodied models, ranking first on multiple core benchmarks.
Generalizable Pick-and-Place
We evaluate generalization under three settings: Basic, Unseen Instructions, and Unseen Objects. Lumo-2 consistently outperforms both π₀.₅ and Fast-WAM baselines across all settings, with particularly large gains on unseen instructions and unseen objects, validating its strong reasoning and generalization capabilities.
Challenging Real-World Manipulation
We test on a diverse suite of real-world tasks organized into three capability dimensions, Lumo-2 achieves the best overall performance across all categories, demonstrating comprehensive embodied intelligence:
- Temporal reasoning: dynamic scene reaction, motion prediction, and multi-phase memory tasks
- Physical understanding: fluid dynamics, gravity interaction, and causal reasoning tasks
- Control complexity: long-horizon multi-step tasks and high-precision dexterous manipulation
Human-to-Robot Transfer
We evaluate emergent cross-domain transfer on two human data sources with distinct view configurations and annotation completeness: multi-view egocentric videos without action annotations, and single head-view VisionPro data with kinematically retargeted action labels. Lumo-2 successfully leverages both data formats to improve robot policy performance on unseen-object robot task settings without requiring any specialized transfer learning mechanisms, confirming the effectiveness of its aligned latent space for cross-domain transfer.
Conclusion & Future Work
Lumo-2 presents a new paradigm for robot learning that unifies latent world modeling and action generation through progressive multi-modality alignment. By replacing explicit reasoning with a compact, physically grounded latent dynamics space, we achieve both strong generalization and efficient inference suitable for real-time control.
The three-stage alignment curriculum addresses a fundamental limitation of existing VLA/WAM approaches: the misalignment between reconstruction-optimized action tokens and downstream control performance. Our results across dozens of real-world tasks confirm that structured latent reasoning and modality alignment are core principles for building capable, scalable, generalist embodied intelligence.
Looking forward, we will continue pushing the scaling frontier by expanding the volume and diversity of training data, scaling up model capacity with broader modality coverage, and exploring more effective data engineering and training paradigms. Beyond scaling, we recognize real-world interactive learning as a core capability for embodied intelligence, and will further investigate deployment-optimized designs and self-evolution mechanisms that enable lifelong adaptation in open scenarios. Ultimately, we strive to build robotic intelligence that delivers meaningful, inclusive value to everyone.
