manipulation

Robotic lab sets up and runs optics experiments on demand (5 minute read)

MIT researchers have built a reconfigurable robotic optics laboratory: a seven-jointed arm picks up lenses and mirrors held in QR-coded 3D-printed housings with magnetic bases, while a Wi-Fi fine-adjustment tool turns their knobs. In a demonstration it assembled a functional tabletop laser cavity through 50 autonomous maneuvers within 30 minutes, and readjusted components to maintain the laser's intensity when parts were moved. The team presents the system at IROS this month.

ForwardDLO: Model-Based Bimanual Shape Matching of Unconstrained Deformable Linear Objects (paper · 15 minute read)

ForwardDLO, a recurrent latent dynamics model for two arms reshaping an unfixated rope, predicts per-segment displacements, making batched evaluation cheap. Its open-loop error on real rope motion is 13% below the strongest baseline, and in a fixed time budget it scores 8 to 22 times more candidate actions than comparably accurate models. On a simulated routing task at 30Hz, that yields 98% success versus at most 30% for baselines.

EgoSteer-RealWorld: 192 hours of teleoperated bimanual manipulation across 193 tasks (dataset · 1 minute read)

EgoSteer-RealWorld, the real-robot dataset used in EgoSteer, holds 54,454 teleoperated episodes (192 hours, 20.75 M frames) of bimanual dexterous manipulation across 193 tasks on a RealMan dual-arm robot, according to its Hugging Face dataset card.

Does Scaling Web-Video Pre-Training Help Real Robots Do Real Work? (13 minute read)

Larger pre-trained video models consistently yield better robot policies, and the trend still holds at the largest size tested, the authors report from a scaling study of Direct Video-Action models pre-trained on web video. More pre-training compute improves policies at every amount of robot data tested, with the largest gains when robot data is scarce. The findings rest on over 200 hours of real-robot evaluation on an industrial bearing-unpacking task.

FoldNet++: a Large-Scale Synthetic Dataset for Robotic T-Shirt Folding and Unfolding (paper · 15 minute read)

FoldNet++ spans 6 robotic embodiments, 1K T-shirts, 1K environmental assets and 120K episodes with rich annotations. It generates physically simulatable T-shirts with annotated semantic keypoints, from which a unified rule-based framework produces demonstrations for each embodiment. The authors report that visuomotor policies trained solely on this synthetic data achieve over 90% end-to-end task success when deployed directly to unseen real-world environments and previously unseen T-shirts.

ARC-Bench: Closed-Loop Replanning Masks Broken Action Ranking in Frozen JEPA World Models (paper · 15 minute read)

Big-D Construction expands FieldAI robot deployments across its project portfolio (4 minute read)

Big-D Construction is expanding its use of FieldAI, with a company executive saying he expects every project to carry some representation of FieldAI tools.

AgiBot says GE-Act 2.0 keeps improving as its training data scales 100× (6 minute read)

GE-Act 2.0 trains every parameter from random initialisation on embodied manipulation data, rather than adding action prediction to a pretrained video generator. AgiBot reports that scaling the corpus from 300 to 30,000 hours lifted zero-shot success on its G1-OP evaluation from 17.1% to 44.1%, and the number of tasks with any success from 39 to 76, with towel-folding and pen-recapping appearing only at the larger scales. It says performance had not saturated at 30,000 hours.

MEMOBench: A Process Level Memory Benchmark for Robotic Manipulation (paper · 15 minute read)

MEMOBench is a process-level memory benchmark for robotic manipulation, testing policies on information that has left the frame.

ContextFlow: In-Context Flow Matching for Robot Manipulation (paper · 15 minute read)

ContextFlow learns robot manipulation from demonstrations held in context, using conditional flow matching rather than the next-token prediction that discretises actions and compounds early errors. Perceiver-style compressors distil visual, proprioceptive and action sequences into compact latents, so several demonstrations can be carried at once. On LIBERO it beats ICRT by 35 percentage points on unseen task configurations and matches a fine-tuned π0 without any fine-tuning of its own.

SmoothRL: Efficient Online RL with Asynchronous Execution (video · 3 min)

Astribot's SmoothRL runs online reinforcement learning while the robot keeps moving, learning from the actions actually executed in the world. The company reports dynamic tossing rising from 39% to 94%, pen capping from 8% to 83% and box opening from 30% to 90%. Astribot calls it the first reinforcement learning validated on real-world, high-dynamic tossing, and has published uncut footage alongside a technical report.

GPT-6 Astra on robot arms (4 minute read)

A follow-up to an earlier test of Claude Fable 5 and 5.1 gives OpenAI's GPT-6 Astra control of the same YAM arms under the same agent policy. Astra placed a block in a bowl in 19 of 20 trials, against 8 for Fable 5.1 and 1 for Fable 5, in 2.5 minutes a trial and at about $0.94 a run. On a puzzle-piece insertion it matched Fable 5.1's 2 successes in 20, stalling at the same final step.

Scaling Bimanual Household Manipulation from 1,500 hours of Demonstrations to On-Policy Corrections (paper · 15 minute read)

The authors release 1,500 hours of bimanual household demonstrations and train XR-2, a vision-language-action model, on them, reporting that success rises steadily with more demonstration data and with DAgger corrections from live human interventions.

Innovation at Agility: Digit Redecorates (video · 3 min)

Agility's Digit rearranges a full room of objects, each a different size, weight and grip. The company presents this as the same generalised manipulation Digit uses for mixed loads and changing layouts on a real warehouse floor. It is posted as a short video demonstration, with no new hardware and no customer deployment attached.

ViFailback's 5,202 episodes of robot failures, now in LeRobot format (dataset · 1 minute read)

ViFailback, the dataset from a CVPR 2026 paper on learning from manipulation failures, now has a LeRobot conversion: 5,202 real ALOHA dual-arm episodes across 100 tasks, 4,545 of them failures, each failure with an annotated keyframe to avoid and one to correct. It is MIT-licensed; the depth maps were left out.

AGIBOT Wins 18 Gold Medals on World Humanoid Robot Games Debut, Tops Gold and Overall Medal Tables (2 minute read)

AGIBOT won 18 gold medals on its World Humanoid Robot Games debut, topping both the gold and overall medal tables.

Keeping the Franka Emika Panda alive: a ROS 2 stack with a reliable position interface (paper · 15 minute read)

The Franka Emika Panda's external position interface has been unreliable for years, and this paper traces the vibrations and protective stops to the timing of the external control loop and sampling jitter rather than to the robot. The open-source stack decouples real-time communication from the ROS 2 control loop and adds rate matching and position-domain reference generation. It is validated across motion planning, compliance control, position-controlled manipulation and haptic teleoperation, on two independent Panda platforms.

ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning (paper · 15 minute read)

ADEPT pretrains a dexterous reinforcement-learning policy on generic object reposing and post-trains task policies from it, then transfers them zero-shot from simulation to a 23-DoF Kuka-Allegro and a 29-DoF Flexiv-Sharpa hand with tactile sensors, solving long-horizon tasks at human-level speed.

LimX Dynamics' TRON 2 patrols an underground utility tunnel on its own (video · 3 min)

LimX Dynamics' TRON 2, fitted with an autonomous mobile manipulation kit, maps and patrols underground utility tunnels, adjusting its posture to reach inspection points at different heights. It detects anomalies, performs light tasks and returns to its charging dock on its own for long unattended shifts. LimX presents the deployment as evidence that partners can build embodied AI applications for city infrastructure and industrial inspection on the platform.

Tactile Sim2Real without Tactile Simulation via Bottlenecked Latent Reconstruction (paper · 15 minute read)

SBLR trains tactile policies on an easily simulated oracle sensor, a point cloud with fingertip forces, and maps the real sensor's embeddings onto it at run time. On peg insertion and gear meshing with GelSight Mini and DIGIT sensors it reaches 85–97.5% zero-shot success without modelling either sensor.

DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation (paper · 15 minute read)

DreamX-Phi 1.0 predicts the frames that follow a prescribed sequence of end-effector poses and gripper states, with per-arm geometric encoding so each arm follows its commanded path, and a depth branch and object masks to keep the scene consistent. It took first place on Track 1 and second on Track 2 of the WorldArena 2.0 Challenge; the model and code are to be released.

Northrop’s robot space mechanic is a new way to keep satellites at work longer (3 minute read)

Northrop Grumman's Mission Extension Vehicle has undocked from an Optus communications satellite after more than a year holding it in position. Four replacement spacecraft launched in July, among them a Mission Robotic Vehicle carrying two robotic arms. The Mission Robotic Vehicle's two arms were developed by DARPA; the other three craft are Mission Extension Pods, smaller modular units.

DecisionFacts publishes a sample of its SO-101 arm teleoperation data (dataset · 1 minute read)

Data vendor DecisionFacts has published an evaluation sample of its SO-101 arm teleoperation data: human-teleoperated manipulation episodes filmed from two cameras at 30 fps, with joint states and actions in LeRobot v3.0 format, and failure and recovery episodes labelled alongside the successes. The sample is licensed CC BY-NC 4.0 for research only; the full catalogue is licensed separately.

Flex-π: A Multi-Stream World-Action Model with Compute Flexibility (paper · 15 minute read)

Flex-π, a 6B-parameter world-action model, is supervised on 3D pointmaps and DINO semantics as well as RGB, all encoded by the same frozen video VAE, and one checkpoint runs on any subset of those streams. The authors report gains of up to 2–7x over the strongest baselines on real bimanual tasks, running faster than π0.5.

TLabel – A unified annotation format and converter for tactile robotics datasets (tool · 3 minute read)

TLabel is a cross-sensor annotation schema for tactile manipulation data, producing 14-dimensional semantic annotations that different sensor types can all emit. Each adapter declares which dimensions it can and cannot annotate, so results from different tactile sensors become comparable rather than sensor-specific. Compliance levels L1 to L4 let each sensor participate at whatever information density it can actually support.

Robot Recycler Salvages Parts From Broken Machines (4 minute read)

Researchers at the Karlsruhe Institute of Technology have built a system that predicts the defect in a broken product and disassembles it while protecting the valuable parts. With more than 4 million industrial robots in use worldwide and over 16 million forecast by 2030, the article argues the industry should prepare to dismantle, recycle and rebuild what it makes.

RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance (paper · 15 minute read)

RynnValue is an open-source value model for robot manipulation trained on temporal distance to the goal, which timestamps supply for free, over more than 7,000 hours of clips. It beats the preference-supervised state of the art on RBM-EVAL-OOD, 0.675 to 0.655, and as a dense reward lifts real-world policy success from 52.5% to 72.5% online.

A ‘5-in-1’ seed-sized surgical robot (4 minute read)

A team at Nanyang Technological University built a 4.4 mm magnetically controlled robot that performs five surgical functions — moving, cutting tissue, releasing drugs, gripping samples and generating heat — and switches between them in under a second. The work is published in Advanced Materials. Led by Associate Professor Lum Guo Zhan, the robot deploys a blade to cut tissue or emits heat to a target area, relevant to the heat-based cancer treatments under study.

CaryX AI releases 10 hours of chest-mounted RGB-D video of two-handed tasks (dataset · 1 minute read)

Everyday Manipulation 3D publishes 1,513 raw RGB-D clips — 10.28 hours across four participants and ten manipulation tasks.

Video to Data: NVIDIA's robot dexterity task library and dataset (dataset · 1 minute read)

NVIDIA has released 1,215 human manipulation demonstrations from the TACO dataset, retargeted to bimanual Sharpa hands, with 121,500 robot episodes generated from them by its Video to Data pipeline. The 924.8 GB set, in Parquet and LeRobot v3 format under CC-BY-4.0, is meant to show what the pipeline outputs and to let users try its grounding step or downstream training.

Beyond VLAs: How World Action Models Reshape Robot Manipulation (5 minute read)

NVIDIA argues that a vision-language backbone learns to describe the world rather than predict how it evolves, which is exactly what a robot needs when a task depends on anticipating change. The post sets out world action models as the replacement for that backbone. NVIDIA researcher Jim Fan set out the shift in his Robotics' End Game talk, later condensed to the line that VLAs are dead.

CaryX AI releases 116 annotated episodes of everyday two-handed tasks (dataset · 1 minute read)

CaryX AI has released 116 episodes of five people doing ten household tasks, recorded with an iPhone Pro's LiDAR, with metric depth, camera pose, MANO hand pose, per-hand object contact, masks and time-aligned language on every frame. It comes in LeRobot v3.0 format under a non-commercial CC-BY-NC-SA-4.0 licence, built for method development rather than broad generalisation claims.

Tactus: Open-Vocabulary Object Recognition from Low-Cost Pressure Arrays (paper · 15 minute read)

Tactus answers text queries about objects from the readings of cheap resistive pressure arrays alone, reaching 0.771 top-1 on the STAG benchmark without a trained classifier head, level with the dataset's supervised CNN. The weights and code are released, and so are the failures.

Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data (paper · 15 minute read)

Ego2Robot turns egocentric videos of people handling objects into robot training data by retargeting their actions, rendering robot arms into the footage and filtering for quality: 18,561 hours across 15 robot morphologies, which the authors call the largest such dataset. Pretraining on it alongside robot data improved out-of-distribution generalisation, including on real robots.

Sharpa's Wave hand served as verification hardware for Gemini Robotics 2 (2 minute read)

Sharpa's Wave dexterous hand, mounted on Apptronik's Apollo 2, served as verification hardware for Google DeepMind's Gemini Robotics 2. The demonstration covered multi-finger tasks including tying trash bags, unscrewing light bulbs and sealing resealable bags. Gemini Robotics 2 extends control from the upper body to the whole humanoid, feet to fingertips, and in a recorded demo guided Apollo across a room to shelve a watering can while avoiding obstacles.

Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them (paper · 15 minute read)

The authors trace the failures of vision-language-action models in contact-rich manipulation to two causes: precision failures, from a mismatch in how flow-matching policies are trained, and force failures, from the structure of force signals. FACT combines a targeted fix for each and reaches a 66% average success rate on five contact-rich tasks, against 41% for the best prior baseline, over almost 2,500 real-world rollouts.