New on arXiv 63RSS 2027 · December 4 deadline
ROBOTICS, IN PERSPECTIVE
From discoveries in the lab to changes on the factory floor. A closer look at robotics today, and what comes next.
About Talos
01 / IN FOCUS
Policy · Mobile robots
At a meeting of the presidential committee on regulatory reform, Korea's Ministry of Land, Infrastructure and Transport set out plans to revise self-driving regulations and to help mobile robots reach the market. The aim is a legal footing for robots in apartment complexes, construction sites, parking lots, airports and railway facilities.
Industry · Robot AI
The company presented Dream-1, a model meant to run on robots from different manufacturers, and launched the Dream API.
Industry · Humanoids
AGCO and Agility Robotics announced an agreement for a humanoid pilot in Hesston, Kansas. Expected to start in 2027, its initial tasks are sorting and stacking parts and placing totes for storage.
RESEARCH IN MOTION / 02
LEARNING THROUGH TASKS
234 tasks. The limits of experience.
Explore this researchSource · arXiv:2610.30876Follow the movement as you scroll, or explore a pose yourself.
RESEARCH DIGEST
MIT’s GelSight Svelte, 2023: a real example of tactile sensing, rather than an experiment newly conducted by the review.MIT / Courtesy of the researchersImage terms
Review · Touch
Touch tells a robot directly what is happening where it makes contact, which makes it central to dexterous manipulation. This review traces the shift from reading tactile signals through contact mechanics to learning representations and policies from data, with imitation learning, reinforcement learning, VLA models and predictive models.
Humanoids · Benchmark
234locomotion and manipulation tasks
LocoManiLab brings together 234 tasks across five kinds of household scenes and 51.5 hours of simulated teleoperation. When twelve policies were compared, success rates fell on tasks and scenes the robots had not seen before.
Space robotics
2 / 200successful captures by a generalist policy
AstroBench is an open benchmark for grabbing free-floating objects inside a spacecraft. A generalist policy caught 2 of 200 stationary targets, while a policy trained only for cubes caught 8 of 10.
Our editors write each digest from the paper's abstract. The numbers and conclusions can be checked in the original paper.
FIELD NOTES / 06 DISCIPLINES
Walking is one challenge. Getting a job done is another. Follow the work of grasping, carrying and adapting to unfamiliar tasks.
From assembling components to sharing a production line: explore how joints, precision and flexible workflows meet on the factory floor.
No two routes are the same. Explore how delivery robots, inspection quadrupeds and drones perceive their surroundings and navigate around obstacles.
Sowing, weeding and harvesting each call for different movements. Explore how robots distinguish crops from weeds, navigate between rows and handle fruit gently as light and soil conditions change.
Further reading
Positioning a weeding tool with a quadruped’s body
Seeing, learning and choosing the next move. Follow transferable robot learning and the systems that automate scientific work.
Exploration and manipulation face unfamiliar conditions in space. From capturing a free-floating object to moving a robotic arm, follow the work beyond Earth.
Further reading
Capturing free-floating objects in space
04 / RESEARCH DESK
Conference opens
Conference on Robot Learning
November 9–12, 2026 · Austin, Texas, USA
Stage 1 extended abstract deadline
Stage 1 extended abstract · Dec 4, 23:59 AoE
KST Dec 5, 20:59 · Final paper submission April 16, 2027 · Conference July 6–11, 2027, Athens, Greece
Conference opens · Seoul
IEEE International Conference on Robotics and Automation
May 24–28, 2027 · COEX, Seoul
Source snapshot: 2026.10.23 · Dates below checked against official guidance on 2026.10.25.
Injects sensor lag, grip slip, unreachable targets and planner-output faults into simulation to identify where failures arise.
Alpha: Systematic Fault Injection and Robustness Evaluation for LLM-Driven Embodied Manipulation
The prototype fixes a weeding tool to Spot and adjusts its position through body motion while the feet stay planted. Indoor and outdoor trials evaluate accuracy and working efficiency as a proof of concept.
Mechanical Precision Weeding with a Quadruped Robot
Tests a world model that estimates garment shape and future state from partial observations during bimanual folding.
In-Context Learning World Models for Robot Dual-Arm Manipulation of Soft Garment with Partial Observations
Selected from papers published by October 23, 2026.
05 / FROM YOUR WORLD
New research, robots at work, and questions that emerge in practice. Share the robotics stories you think deserve a closer look.
