IROS 2026 Pittsburgh — Navigation and Manipulation Converge
IROS 2026 (Sept 27 - Oct 1, Pittsburgh) continues industry dominance with autonomous navigation and robot manipulation as the top themes. The boundary between navigation and manipulation is blurring — robots increasingly need both skills simultaneously in complex environments.
Key Trends
Navigation Innovations
-
Hybrid Motion Planning with Deep RL: Combines classical planning (A*, RRT) with learned local components for better generalization in dynamic environments.
-
Human-like Navigation using VLM Reasoning: Robots learn social norms and human-like behavior through Vision-Language Model reasoning, not just collision avoidance.
-
Decentralized Multi-Robot Coordination: Multi-robot navigation in GPS-denied, communication-limited environments using hierarchical topological sharing.
Manipulation Advances
-
DexUMI: Human Hand as Universal Interface: Using hand tracking (MediaPipe) directly as teleoperation interface for collecting dexterous manipulation data.
-
Sim-to-Real for Long-Horizon Pick-and-Place: Fully autonomous long-horizon tasks with robust perception pipeline, adaptive grasp planning, and error recovery.
-
Dexterous Manipulation via Imitation Learning: Comprehensive survey of state-of-the-art methods covering data collection, learning approaches, and sim-to-real transfer.
Notable Papers
Hybrid Motion Planning (arXiv:2512.24651)
Classical planners (A*, RRT) ensure completeness but are slow. RL planners react fast but lack global reasoning. This paper combines both:
- Global planner: ensures robot reaches destination
- RL local planner: handles dynamic obstacles in real-time
- Smart switching based on uncertainty estimation
Result: 15% improvement in success rate vs pure RL, 25% faster than pure classical planning.
Human-like Navigation (arXiv:2509.21189)
Robots typically treat people as obstacles. This paper enables social navigation:
- Vision-Language Model (VLM) reasoning about context
- Robot understands doors, corridors, crowded areas
- Chooses appropriate behavior (wait, yield, detour) vs just avoiding
Practical impact: Essential for service robots in hospitals, hotels, shopping centers.
DexUMI Paper (arXiv:2505.21864)
Problem: Collecting dexterous manipulation data requires skilled operators, is expensive and slow.
Solution: Use human hand tracking directly as manipulation interface
- Operator thaws hand, system captures via camera
- Retargets to any robot hand automatically
- 86% average success rate, transfers across platforms
Takeaway: Faster, cheaper, more natural data collection than traditional teleoperation.
Convergence: Mobile Manipulation
Navigation and manipulation blur together in IROS 2026. Robots must move to locations AND interact with environment. This mobile manipulation is the real-world requirement for service robots.
Foundation Models Everywhere
From navigation (VLM reasoning), manipulation (VLA models), perception (foundation detection models), to multi-robot (LLM task planning) — foundation models appear throughout. This represents a paradigm shift.
Safety is Mandatory
Papers on safe RL, formal verification, and human-aware planning increased significantly. Safety analysis is now required by reviewers — no longer optional.
For Vietnamese Engineers
Warehouse automation, e-commerce logistics demand both navigation (autonomous movement) and manipulation (picking items). IROS papers directly address this convergence.
Action items:
- Study hybrid motion planning for warehouse AMRs
- Implement imitation learning pipelines for manipulation
- Use foundation models for rapid prototyping
- Add safety constraints to RL policies before deployment
