Robotics: January 2026 Week 3
Jan 15 – Jan 21, 2026 · 71 papers analyzed · 3 breakthroughs
Summary
Week 3 (Jan 15-21): 3 breakthroughs from 71 papers. (1) 2601.09988 (UMI-FT) enables in-the-wild compliant manipulation via finger-level F/T sensing; (2) 2601.10930 introduces hierarchical RL-MPC with contact-intention interface for dexterous manipulation; (3) 2601.12428 (ReWorld) tackles the physics uncanny valley with multi-dimensional reward modeling. Dexterous manipulation and contact-rich control dominate.
Key Takeaway
The field is moving from open-loop manipulation to contact-aware, physics-grounded policies.
Breakthroughs (3)
1. In-the-Wild Compliant Manipulation with UMI-FT
Why Novel: First system providing finger-level force/torque sensing during in-the-wild demonstrations, enabling capture of both external contact and internal grasp forces.
Key Innovations:
- CoinFT sensors on each finger for F/T during data collection
- Adaptive Compliance Policy learns force modulation from demonstrations
- Low-cost, portable system deployable outside lab settings
Evidence:
- — CoinFT sensor integration and calibration
- — Adaptive Compliance Policy architecture
- — Force-sensitive task completion rates
Impact: Unlocks contact-aware policy learning from natural demonstrations, critical for delicate manipulation.
2. Where to Touch, How to Contact: Hierarchical RL-MPC Framework for Geometry-Aware Long-Horizon Dexterous Manipulation
Why Novel: Introduces contact-intention interface that decomposes dexterous manipulation into geometric planning (where to touch) and contact dynamics (how to contact).
Key Innovations:
- High-level RL predicts contact points and post-contact subgoals
- Low-level MPC handles nonsmooth contact dynamics
- Geometry-aware representations enable long-horizon planning
Evidence:
- — Contact-intention interface formulation
- — Hierarchical RL-MPC architecture
- — Long-horizon manipulation sequences
Impact: Bridges the gap between task-level planning and contact-level execution in dexterous hands.
3. ReWorld: Multi-Dimensional Reward Modeling for Embodied World Models
Why Novel: Tackles the 'physics uncanny valley' in video world models by introducing HERO, a four-headed reward model trained on 4D Embodied Preference Dataset.
Key Innovations:
- 4D Embodied Preference Dataset with multi-dimensional annotations
- HERO decouples physics, consistency, grounding, and task rewards
- HERO-FPO enables tractable flow-based policy optimization
Evidence:
- — 4D preference dataset collection methodology
- — HERO multi-headed reward architecture
- — Physics realism and task success improvements
Impact: Provides principled approach to training physically grounded world models for robot learning.
Trends
Contact-rich manipulation seeing increased attention with force sensing and contact planning
World models being refined with physics-aware reward functions
Bimanual and dual-arm systems gaining traction for complex assembly tasks
Notable Papers (4)
1. FastStair: Learning to Run Up Stairs with Humanoid Robots
GPU-parallel DCM planner with multi-stage RL for agile stair climbing.
2. A3D: Adaptive Affordance Assembly with Dual-Arm Manipulation
Per-point affordances for furniture assembly with dynamic strategy adjustment.
3. BiKC+: Bimanual Hierarchical Imitation with Keypose-Conditioned Consistency Policies
Keypose predictor + consistency model for coordinated bimanual manipulation.
4. Learning Legged MPC with Smooth Neural Surrogates
Lipschitz-controlled MLP surrogates enable stable gradient-based MPC for legged robots.
Honorable Mentions
- X-Distill: Cross-Architecture Vision Distillation for Visuomotor Learning ()
- ACoT-VLA: Action Chain-of-Thought for Vision-Language-Action Models ()