Robotics: February 2026 Week 7

1. LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion

arXiv:2602.12215

Why Novel: First robot foundation model to jointly learn policy, dynamics, and visual forecasting from heterogeneous embodied data (real robots, simulation, humans) at 1B+ scale via semantically structured latent dynamics and role-aware supervision.

Key Innovations:

• Universal embodied data ingestion from EI-30K spanning real robots, simulation, and human demonstrations
• Semantically structured latent space with multi-modal diffusion transformer for joint policy/dynamics/forecasting
• Role-aware supervision handling mixed-quality data with embodiment-specific objectives
• Strong cross-embodiment transfer to contact-rich and dexterous manipulation tasks

Evidence:

• fig:2 — LDA architecture showing latent dynamics and multi-modal transformer
• tab:1 — Performance across simulation benchmarks showing scaling benefits
• tab:2 — Real-world transfer results on contact-rich manipulation
• fig:5 — Scaling curves demonstrating data efficiency gains from latent dynamics

Impact: Provides a practical pathway to scalable robot pretraining that leverages the massive diversity of existing embodied data while learning transferable dynamics rather than just mimicking actions.

2. Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment

arXiv:2602.12281

Why Novel: First systematic demonstration that test-time verification scaling outperforms policy scaling for VLA alignment, achieving 45% real-world improvement via a 1B contrastive verifier (CoVer) with hierarchical language-action optimization.

Key Innovations:

• Contrastive verifier (CoVer) trained offline to assess instruction-action alignment at test time
• Boot-time compute allocation for instruction rephrasing and action probing based on semantic alignment
• Hierarchical optimization pipeline selecting both instructions and actions via verifier scores
• 22% in-distribution, 13% OOD, and 45% real-world gains on SIMPLER with lower compute than policy scaling

Evidence:

• fig:2 — CoVer architecture and hierarchical optimization pipeline
• tab:1 — SIMPLER benchmark showing verification scaling beats policy scaling
• tab:2 — PolaRiS results with 14% task progress and 9% success rate gains
• fig:5 — Scaling curves comparing verification vs policy compute allocation

Impact: Shifts the efficiency frontier for VLA systems by demonstrating that runtime reasoning can be more effective than pretraining scale, suggesting a fundamental change in how robotics should allocate compute.

3. GigaBrain-0.5M*: a VLA That Learns From World Model-Based Reinforcement Learning

arXiv:2602.12099

Why Novel: First VLA to condition policy predictions on future-state and value estimates from a pretrained world model, enabling self-improvement via human-in-the-loop rollouts and continual RAMP (Reinforcement Augmented Model Prediction) training.

Key Innovations:

• RAMP framework conditioning VLA on world model future-state and value predictions
• Self-improvement loop via human-in-the-loop rollouts and continual training
• Strong long-horizon manipulation with cross-task generalization
• State-of-the-art on internal tasks and RoboChallenge benchmarks

Evidence:

• fig:2 — GigaBrain architecture showing world model conditioning
• tab:1 — RoboChallenge benchmark results achieving SOTA
• fig:5 — Long-horizon task performance with world model vs without
• fig:8 — Real-world deployment across diverse manipulation scenarios

Impact: Demonstrates that world model conditioning can provide VLAs with the foresight needed for reliable long-horizon manipulation, offering a scalable path to anticipatory robot control.

1. HoloBrain-0: Comprehensive VLA Framework with Embodiment Priors

arXiv:2602.12062

Cross-embodiment VLA grounding policies in URDF and camera parameters with RoboOrchard infrastructure achieving SOTA on multiple benchmarks.

2. JEPA-VLA: Video Predictive Embedding is Needed for VLA Models

arXiv:2602.11832

Shows V-JEPA2 video embeddings capture task-relevant dynamics, improving VLA sample efficiency and generalization.

3. VLAW: Iterative Co-Improvement of VLA Policy and World Model

arXiv:2602.12063

Real-robot framework co-improving VLA and world model on DROID platform with synthetic trajectory generation.

4. Robot-DIFT: Distilling Diffusion Features for Geometrically Consistent Visuomotor Control

arXiv:2602.11934

Manifold distillation from diffusion models yields geometry-preserving backbone for precision manipulation.

5. FAST: General Humanoid Whole-Body Control via Pretraining and Fast Adaptation

arXiv:2602.11929

CoM-aware pretrained controller with Parseval-regularized residual policy for rapid humanoid adaptation.

6. Xiaomi-Robotics-0: Open-Sourced VLA Model with Real-Time Execution

arXiv:2602.12684

Industry-grade open VLA optimized for real-time deployment with cross-embodiment pretraining.

7. EasyMimic: Low-Cost Robot Imitation Learning from Human Videos

arXiv:2602.11464

Sub-$300 setup enabling manipulation learning from consumer RGB videos via action retargeting.