RSS 2026: The Highest Quality Robotics Conference Comes to Sydney
Robotics: Science and Systems (RSS) 2026 — the 22nd edition — will be held from July 13-17 at University of Technology Sydney, Australia. While ICRA is the largest conference, RSS is the highest quality in robotics: single-track format ensures all attendees see the same presentations, acceptance rate is much lower than ICRA, and every paper undergoes strict peer review.
This year, RSS comes to the Asia-Pacific region for the first time in years — an excellent opportunity for Vietnam's robotics community to access a top-tier conference without flying to the US or Europe. Workshops run on Monday July 13 (half-day morning) and Friday July 17 (two half-day events), with main conference sessions from Tuesday through Thursday.
This article previews the most noteworthy research directions you can expect at RSS 2026, based on trends from ICRA 2026 and recent papers on arXiv.
Research Direction 1: Next-Generation Vision-Language-Action Models
RSS has always been where the biggest breakthroughs in robot learning debut. In 2024, Octo launched at RSS and became the baseline for all subsequent VLA research. This year, the next generation of VLA models will dominate.
From pi-0 to pi-0.5: Open-World Generalization
pi-0 from Physical Intelligence was already a game-changer by proving one model can control multiple robot configurations on 68 different tasks. But pi-0.5 takes everything to the next level: robots can clean kitchens and bedrooms in completely new homes they've never seen before.
Key innovation is co-training on heterogeneous data sources: multiple robots, semantic predictions, web data, and object detections. RSS 2026 will feature many follow-up papers in this direction — especially on data scaling laws and transfer efficiency.
GR00T N1: NVIDIA's Open-Source Push
GR00T N1 with dual-system architecture (vision-language reasoning + diffusion transformer actions) has been fully open-sourced. The RSS community, which values reproducibility and open science, will see many papers building on top of GR00T — from fine-tuning for specific tasks to scaling across new embodiments.
Prediction: RSS 2026 will have at least 5-10 papers using GR00T N1 or pi-0 as baseline, showing consolidation of VLA research around a few core foundation models.
Research Direction 2: More Sophisticated Sim-to-Real Transfer
RSS has witnessed many sim-to-real milestones, from DrEureka (LLM-guided reward design) to Humanoid-Gym (zero-shot locomotion transfer). In 2026, research is shifting in two new directions.
World Models for Sim-to-Real
World Models Can Leverage Human Videos for Dexterous Manipulation — DexWM — is particularly interesting. Instead of using only simulation data, DexWM trains a world model from human videos then optimizes actions via MPC framework. Result: 83% success rate on object grasping with real Franka Panda + Allegro gripper — zero-shot from human video data.
This direction is very promising because human video data is nearly unlimited on the Internet, while robot demonstration data remains expensive and scarce. RSS 2026 will feature many papers exploring human-to-robot transfer via world models.
Automated Real-to-Sim Calibration
The paper Sim-to-Real RL for Vision-Based Dexterous Manipulation on Humanoids introduced automated real-to-sim tuning modules — automatically calibrating simulation parameters to match the real world. This is an important shift: instead of manually tuning domain randomization ranges, we're automating the process.
Prediction: "Automated sim gap reduction" will be a keyword at many presentations. Tools like NVIDIA Isaac Lab will integrate automated calibration, transforming sim-to-real from art to engineering.
Research Direction 3: Advanced Dexterous Manipulation
RSS is particularly strong in manipulation research — from DexGraspNet to AnyRotate, many breakthroughs in this area debut at RSS. This year, three directions will stand out.
In-Hand Manipulation with Tactile Feedback
In-Hand Manipulation of Articulated Tools with Dexterous Robot Hands with Sim-to-Real Transfer takes manipulation to the next level: not just grasping objects but manipulating articulated tools — like scissors, pliers — with robot hands. The pipeline combines simulation-trained base policy with sensor-driven refinement from hardware demonstrations, using cross-attention for tactile and force feedback.
This research direction is very relevant for manufacturing in Vietnam — where robots need to handle many different tools in assembly lines.
Zero-Shot Grasping with Tactile Simulation
Closing the Reality Gap: Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping is the first time controllable grasping on multi-finger hands was trained entirely in simulation then transferred zero-shot to real hardware. The key contribution is computationally fast tactile simulation computing distances between dense virtual tactile units and objects via parallel forward kinematics — providing high-rate, high-resolution touch signals for RL.
3D World Models for Multi-Material Manipulation
ParticleFormer: A 3D Point Cloud World Model for Multi-Object, Multi-Material Robotic Manipulation solves an extremely difficult problem: manipulation with multiple different materials (rigid, deformable, flexible) in the same scene. Transformer-based world model trains directly from real-world perception data, capturing fine-grained material interactions.
Prediction: Tactile-aware manipulation will be a major theme at RSS 2026. Tactile sensor prices are dropping (GelSight Mini ~$300), making this research direction more accessible.
Research Direction 4: Navigation with Foundation Models
Embodied Navigation Foundation Model
NavFoM: Embodied Navigation Foundation Model is a noteworthy paper: a navigation model achieving SOTA on diverse benchmarks for multiple embodiments (humanoids, quadrupeds, drones, wheeled robots) without task-specific fine-tuning. Real-world experiments succeed on multiple platforms.
Improved Vision-Language Navigation
VL-Nav: Neuro-Symbolic Vision-Language Navigation combines symbolic 3D scene graphs with VLM reasoning — a hybrid approach between classical planning and neural reasoning. Results are more robust than pure neural approaches thanks to explicit spatial reasoning capability.
Insight for Vietnamese Engineers: If you're building delivery robots or service robots, navigation + natural language understanding is the deciding combo. Read more on SLAM and Navigation for Autonomous Robots to solidify fundamentals.
Noteworthy Workshops
Semantic Reasoning and Goal Understanding (SemRob)
Workshop SemRob in its 3rd iteration focuses on semantic-level understanding for robot task planning — combining knowledge representation with learning-based approaches. Relevant for AI-powered robotics applications.
RSS Pioneers
RSS Pioneers program is for young researchers (PhD students, early postdocs) — an excellent networking opportunity. If you're doing a PhD in robotics in Vietnam or abroad, this is a program worth applying to.
Preparing for RSS 2026
If You Attend In-Person
- Book early: Sydney in July is winter but pleasant (10-17 degrees C). Accommodation near UTS should be booked 2-3 months in advance
- Register for workshops separately: Workshops often fill up early, register as soon as registration opens
- Poster sessions are gold: RSS poster sessions have high quality — prepare questions in advance for papers you're interested in
If You Follow Remotely
- Follow Twitter/X: @RoboticsSciSys will post highlights in real-time
- Read proceedings after the conference: RSS papers are open-access, download free from roboticsconference.org
- Watch recorded talks: RSS usually uploads talk recordings after the conference
RSS vs ICRA: What's the Key Difference?
| ICRA | RSS | |
|---|---|---|
| Format | Multi-track (parallel sessions) | Single-track (everyone sees everything) |
| Papers | 3000-4000 submissions, ~40% acceptance | 500-700 submissions, ~25% acceptance |
| Strengths | Broad coverage, industrial presence | Depth, theoretical rigor |
| Workshops | 80-100 workshops | 15-20 workshops (more selective) |
RSS acceptance rate is significantly lower, meaning every paper undergoes strict review. If you only have time to read papers from one conference, RSS papers give the highest ROI.
