Reading Modes
Papers & Evaluations with links to supporting source material.
556 selected reportsLast 30 days 491 recordsAll coverage 2,632 records
This research paper introduces a robot exposure index to assess how well robots can perform job tasks today. It finds that robots can handle 74% of physical tasks in the US, making up 34% of working hours, but are cost-competitive for only 0.3% of tasks. The study highlights barriers to adoption, including environmental adaptability and human preferences, and suggests that physical work may be automated first where robots already have a foothold.
This paper identifies modality shortcuts in vision-language-action (VLA) policies, where policies rely on superficial cues like visual, lexical, or motor signals rather than task-relevant evidence. The authors propose Perturbot, a method that introduces task-preserving wrist-view perturbations, enriched instructions with decision-relevant captions, and relabeled trajectory segments to reduce reliance on these shortcuts. They also introduce GroundFscore, an offline score to evaluate how much a policy depends on shortcuts. The work provides a framework for training and evaluating VLA policies that disentangle decisions from shortcut priors while preserving responsiveness to task-relevant evidence.
MIT researchers have developed a new autonomous navigation system called SANDO that guarantees collision-free flight for UAVs in dynamic, unknown environments. The system uses a time-sensitive safety corridor and mathematical proofs to ensure safety, even when obstacles are moving unpredictably. SANDO has been tested in simulations and real-world flights, successfully avoiding collisions in all scenarios.
This paper presents a collection of recent research papers and datasets focused on vision-language-action (VLA) models in robotics. Topics include recursive video in-context learning, simulation-guided world models, closed-loop benchmarks for driving, and methods for improving generalization and safety in VLA systems. The work highlights challenges such as task success vs. instruction following, the grounding gap, and the need for efficient and reliable adaptation in real-world scenarios.
NVIDIA has released a policy model trained on LIBERO-10, part of the Cosmos Cookbook recipe. The model, trained with 379 real demonstrations and 1137 duplicates, achieves 94.2% success in standard scenes and 85.3% in perturbed scenes. The model is available under the NVIDIA Open Model License.
TouchScale is a 500-hour human vision-touch dataset collected using a unified sensor setup, enabling scaling benefits in both perception and robotic manipulation. The dataset improves robot task success rates through mid-training with human data, demonstrating the potential of large-scale human experience in training embodied agents.
This paper introduces ChunkVLA-AM, a reproducible deployment pipeline for vision-language-action (VLA) models on a FAIRINO FR3 robot in an additive manufacturing (AM) workcell. The system converts real-world demonstrations into OpenVLA-compatible datasets and uses a cloud-edge architecture for inference. It achieves 92.9% success in transferring objects between points in a physical test, with failures attributed to insufficient release-height control. The study also evaluates the model's performance under varying illumination conditions, finding that prediction errors are lowest within a specific luminance range.
Rho is a new foundation model for vision-language-action (VLA) models in robotics, designed to improve adaptation to specific robots and tasks. The paper introduces a training recipe that separates embodiment adaptation from task adaptation, reducing the amount of task-specific data needed for fine-tuning. Rho variants are pre-trained on different robots and show strong performance on physical tasks, outperforming existing models in several benchmarks. The model also supports online adaptation through a latent policy that learns from human corrections.
This paper presents a systematic analysis of Vision-Language-Action (VLA) research, focusing on datasets, benchmarks, and data engines. It identifies key challenges in data fidelity, scalability, and evaluation, and proposes a structured approach to address these issues. The authors release an open-source repository to support the community.
Carnegie Mellon University researchers have developed a robotic fingertip that enables robots to handle delicate tasks, such as picking up fragile or tiny objects like pills and potato chips. The fingertip, which attaches to the DeltaHands device, uses a strain gauge and contact microphone to measure pressure and capture microvibrations, allowing the robot to sense and manipulate objects with greater precision. The low-cost, 3D-printable design offers versatility for a range of tasks, from counting paper cups to identifying material compositions through vibrations. The work was presented at the 2026 Intelligent Robots and Systems Conference.
This paper presents LeaP, a learnable source prior for generative robot policies, which replaces fixed standard Gaussian distributions with state-dependent random initialization. LeaP jointly learns the mean and variance of an initial Gaussian distribution, improving action generation in robotic tasks. The method achieves an average success rate of 81.6% in simulation and 80.0% on real robotic arms, outperforming existing baselines like A2A and VITA. The paper is accepted at CoRL 2026, and the code is open-sourced.
Sanctuary AI introduces TIGER, a reinforcement learning framework that uses imitation policies as dense reward signals to improve efficiency and safety. TIGER reduces interaction steps and safety resets on both simulated and real-world tasks, demonstrating its potential for deployment in industrial robotics.
Professor Shi Qing's team from Beijing Institute of Technology was inspired by the high-speed running mechanism of the elephant shrew. They designed a micro quadruped robot named FLEXOR equipped with a dual-joint coupled spine. Through dynamic spine-leg coordination, the robot achieved a 31.6% increase in speed and a 32.2% reduction in energy consumption without upgrading its motors. The study reveals the critical role of the timing coordination between the spine and legs in locomotion performance and verifies the universality of this mechanism across different robot morphologies, providing new insights for highly mobile embodied intelligent robots.
Researchers from the National University of Singapore have developed a continuous artificial muscle inspired by the neuromuscular control of an octopus. This design allows for precise, localized control of soft robotic arms without the need for multiple actuators. The system enables compact, lightweight robotic arms capable of complex movements and tasks such as sampling and object retrieval. The approach could have applications in confined spaces, aerial deployment, and industrial settings.
Researchers at Tsinghua University have developed a new control system for humanoid robots that enables coordinated soccer-playing behavior using onboard vision. The system was tested on a real robot and achieved an 80-90% success rate in scoring goals from the front half of the field, with a 64% reduction in response time compared to a rule-based system. The study was published in Science Robotics and demonstrates a significant advancement in embodied AI and robot coordination.
Robotic Systems Lab@leggedroboticsSignalEditorial score7878
MIT engineers have developed a soft, paper-thin robot that swims through water using a single layer of genetically engineered muscle cells. The robot's 'fins' are covered with live muscle cells that twitch in response to light, allowing it to flap and navigate through a watery maze. The study demonstrates the potential of biohybrid robots for delicate aquatic tasks, with the robot achieving a swimming speed of about four times its body length per minute.
Researchers from North Carolina State University have developed a soft, non-electronic robot that jumps continuously using infrared light. The 'soft ring jumper' uses elastic energy stored through torsion, with a V-shaped aluminum stopper controlling the direction of movement. By adjusting the geometry of the V-stop, the robot can crawl, jump forward, or leap vertically. The design enables autonomous jumping without external resets, opening possibilities for use in environmental navigation and unstructured terrain.
Carnegie Mellon University researchers have developed LAMP, a system that allows multiple robots to work together to move objects through crowded spaces. LAMP combines learned models with search-based planning to enable efficient navigation and coordination, with successful testing in complex environments and a demonstration at the 2026 IROS conference.
Cornell researchers have developed microscopic robots that can sense temperature and work together to alter their environment by pumping liquid between hot and cold regions. Inspired by natural systems like fireflies and cilia, these robots use synchronized motion and electronic communication to achieve collective behavior, marking a significant step toward advanced micro-robotics with potential applications in medical and agricultural settings.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).