Robotics Briefing /
Robots learn to handle the messy parts.
A folded shirt can look wrong even when a robot finishes the task. Sand can give way under a perfectly planned step. Today’s research puts those physical details into robot training and evaluation, while new deployments show how automation reaches existing workplaces.
Your quick takeaways
- FolDeX measures real garment folding by completion, neatness, and speed. Recovery examples helped one policy reach 95% average success across four garment categories.
- Humanoid training that models sinking and sliding feet transfers to loose ground. DUET-DINO improves planning with two camera views, with hardware results still far below its simulation scores.
- General Robotics is automating parts of robot engineering. Nauticus is selling supervised navigation assistance for existing underwater vehicles, and MBody AI reports a paid deployment after a hospitality pilot.
Papers
FolDeX: measuring whether a robot folds well
Chenhuan Liu and colleagues • Submitted September 9 • Preprint; physical-robot benchmark
Cloth changes shape with every grasp, so small mistakes can accumulate over a long folding sequence. FolDeX introduces a physical benchmark for garment manipulation, with standardized starting conditions and scores for completion, final neatness, and execution time. The authors describe more than 2,000 hours of robot experience across over 20 tasks and 10 robot embodiments.
A single π0 policy trained with demonstrations and recovery examples achieved 95% average success across shirts, skirts, pants, and towels. The protocol generally uses 30 physical trials per garment category, and success requires completing the full sequence without human intervention. Neatness remains a separate measure: a completed fold can still be poor.
This makes recovery from mistakes and finished quality visible in evaluation. The experiments remain preliminary and costly, use no explicit touch sensing, and show that mixing data across robots can hurt performance. The linked FoldChallenge platform hosts evaluation; access to the complete data collection was not independently verified.
Humanoid locomotion learns what happens when the ground gives way
Junnosuke Kamohara and colleagues • Submitted September 9 • Preprint; simulation and Unitree G1 tests
Training on rigid ground misses the resistance a foot encounters as it sinks into sand or drags through gravel. This method models those forces during reinforcement learning, then trains a controller to infer changing terrain stiffness and adjust its movement.
On physical dry sand, the authors report stable locomotion at commanded speeds up to 2.5 meters per second; the rigid-ground baseline failed above 1.5 meters per second. They also tested basalt and beach terrain. Separately, a deep-sand simulation comparison achieved 85% success with their contact model, versus 3% and 8% for two simplified models.
The practical implication is that more realistic contact physics can teach useful behavior before hardware trials. Those simulation percentages are not outdoor reliability rates, and the physical experiments cover a limited selection of surfaces. The project page includes demonstrations and marks code as coming soon.
DUET-DINO: two camera views improve planning with a learned world model
Nisarga Nilavadi and colleagues • Submitted September 9 • Preprint; simulation and physical-arm tests
A wrist camera sees the grasp closely; a side camera sees the surrounding workspace. DUET-DINO lets predictions from both views inform each other as a learned world model estimates the consequences of possible robot actions. The planner uses those predictions to choose position, orientation, and gripper movements.
The paper reports 92% success on reaching in simulation. In a separate physical evaluation of difficult reaching poses and rotations, it achieved 26.7% across 30 trials on three tasks, compared with 16.7% for independently modeled camera views. Hardware used a smaller planning budget because computation and latency constrained execution.
Combining local and wider scene information is promising for precise manipulation, but this remains an early planning result. The authors report 15–17 seconds per planning step in their experimental setup. The project lists code and checkpoints as forthcoming; it does not yet provide those releases.
News
General Robotics connects engineering agents to physical tests
Announced September 9 • GRID Auto-Engineering; company experiments
General Robotics introduced Auto-Engineering workflows that connect hardware integration, simulation, skill development, and deployment feedback. Its laboratory examples include transferring a pouring behavior between robot arms and building a simulation that couples a fluid solver with rigid-body physics.
The company reports about four hours from a fresh Flexiv setup to its first working skill, with later skills taking as little as 10–15 minutes using retained engineering work. These are company examples, without an independent comparison across deployments. The useful prospect is reusing tested integrations and fixes so each new task requires less repeated engineering.
Nauticus brings pilot assistance to existing underwater vehicles
September 9 release details • Commercial software; company claims
Nauticus provided further details on the commercial ToolKITT release first announced during its recent earnings call. The available Pilot Assist module supports position holding, current compensation, and waypoint navigation on remotely operated underwater vehicles. A human continues to supervise the mission.
The integration retains the vehicle manufacturer’s manual controls and diagnostics, with an onboard sensor and topside computing package. This offers a route to upgrading existing fleets. Nauticus reports efficiency gains above 20% in customer operations, but the announcement provides no detailed measurement protocol or independent validation.
MBody AI reports paid service following its Mohegan Sun pilot
Announced September 9 • Hospitality deployment; company and customer statements
MBody AI says its pilot robots remain in service at Mohegan Sun under a paid commercial agreement. The pilot covered the gaming floor and conference center during day and evening shifts, coordinated through its fleet-management platform.
A customer continuing to pay after a pilot is a useful deployment milestone. A larger, multi-year expansion remains under negotiation. The release includes a named customer testimonial, but no robot count, intervention rate, or detailed productivity results to establish how broadly the outcome could transfer.
Highlights
Light REACT explores movement after hardware disruption
Published September 9 • Company technical post and demonstrations
Light Origins describes a humanoid controller that uses recent movement history to adapt to disturbances and impaired joints. Its training combines specialist controllers into a single model, then rewards upright movement when possible while retaining crawling as a fallback.
The demonstrations explore recovery and continued mobility when normal walking becomes difficult. That matters for reducing prolonged stoppages, but the company post does not establish reliability across an operating fleet. Its simulated damage coverage and selected physical demonstrations should be judged separately.
Read this first
Read FolDeX’s evaluation protocol and recovery-data results. Separating completion, neatness, and speed makes it easier to judge whether an impressive robot demonstration would produce a useful finished result.
Read the paper