Robotics Briefing /
Teaching robots how much force to use.
Peeling a cucumber and picking a ripe berry both require careful contact. Today’s research tests how robots sense pressure, handle delicate objects, and reuse computation during repetitive work. A portable rescue robot and new funding for robot chips show two other routes toward machines that work beyond the lab.
Your quick takeaways
- TAO-Force combines wrist sensing with fast force control. It improved success when researchers changed tool or object geometry, though difficult tasks still caused failures.
- CLASP completed 23 of 25 blueberry-cluster harvesting attempts at one farm. Speed, visibility, and fruit quality remain constraints.
- D-Robotics announced $400 million for robot chips and software. SPROUT offers a public hardware design for exploring confined spaces.
Papers
TAO-Force: helping robots feel and respond to contact
Bohan Gan and colleagues, China Mobile (Hangzhou) Information Technology • Submitted September 16 • Preprint; physical robot tests
A robot peeling a cucumber needs to adjust pressure as the blade moves. TAO-Force feeds wrist force measurements into a vision-language-action model, which predicts movements and desired forces. A separate controller makes rapid corrections when the tool touches an object.
The model runs five times per second; the force controller runs 1,000 times per second. This lets the system respond to contact between the model’s decisions.
Across four physical tasks, success averaged 80% under familiar conditions. After changes such as raising a soap bottle or replacing a knife, it reached 64.2%, compared with 29.2% for π0.5 and 20% for GR00T N1.7. Each task and condition used 30 trials.
The comparison models lacked force input, so the gains reflect the combined sensing and control system. Cutting with the replacement knife still succeeded only 46.7% of the time. These results support combining learned actions with fast physical feedback, while leaving substantial room to improve reliability.
CLASP: picking ripe blueberries a cluster at a time
Yixuan Xia, Yilin Cai and colleagues, Georgia Tech, University of Georgia and University of Florida • Submitted September 16 • Preprint; field tests
Blueberry clusters contain ripe and unripe fruit close together. CLASP wraps two soft rolling bands around a cluster and controls their force to detach ripe berries while leaving harder-to-detach fruit behind. Cameras guide the arm toward the cluster.
At one research farm, the system completed 23 of 25 autonomous harvesting attempts. Both failures occurred before the gripper reached the fruit. Operators drove the platform between stationary harvesting locations, so the study does not establish autonomous operation across an entire farm.
A separate test, with a person aligning the gripper, reached 32 berries per minute versus 55 for hand picking. That measures the gripper’s speed, excluding autonomous approach time.
Fruit quality also needs closer testing: robot-picked berries measured 15% lower in firmness, but waited longer before measurement. Matched timing would help separate handling damage from storage effects. The approach could simplify picking within crowded clusters; larger trials must establish speed and quality alongside successful grasps.
rMuscle: reusing computation during repetitive robot work
Kaijun Zhou and colleagues, Shanghai Jiao Tong University • Submitted September 16 • Preprint; inference benchmarks and physical tests
A robot repeating a packing task sees similar scenes and makes similar movements. rMuscle stores useful internal computations from earlier executions, then reuses them when the current situation matches. It saves work in both visual processing and action generation.
For the π0.5 model, inference time fell from 46.0 to 35.5 milliseconds on an RTX 4090, and from 106.2 to 74.1 milliseconds on Jetson Thor. These are model-response measurements; factory throughput depends on physical motion and other delays too.
In physical trials, the cached model matched the original’s observed success: 76% for cooperative bottle placement and 84% for conveyor packing, with 50 trials per method per task.
This offers a way to reduce computation for familiar work without retraining a larger model. The tradeoff is extra memory and a collection of representative prior executions. Some simulation tasks lost accuracy when similar-looking situations called for different actions, underscoring the need for careful fallback rules.
News
D-Robotics raises $400 million for robot computing
D-Robotics • Announced September 17 • Company funding announcement
D-Robotics announced a $400 million Series C to expand its Sunrise chip family and build software covering the robot development process. Its products serve categories including humanoids, quadrupeds, and logistics robots.
The company says cumulative Sunrise chip shipments have exceeded eight million. That figure covers the chip family across applications; it does not measure humanoid deployments. It also reports more than 20 embodied AI customers for the S600 chip, which launched in November 2025.
The practical opportunity is less integration work between robot hardware, models, and development tools. Shared infrastructure could help manufacturers bring additional robot designs to market.
The announcement provides funding and adoption claims, with no independent comparison of energy use, inference performance, or deployment reliability. Those measures will help establish what the expanded platform delivers.
Highlights
SPROUT: a soft robot that grows into narrow spaces
MIT Lincoln Laboratory and University of Notre Dame • Submitted September 15; listed September 17 • Preprint; public hardware and software
SPROUT extends a pressurized fabric tube from its tip, letting an operator steer it through gaps in rubble. A battery and a standard firefighter air tank make the system portable, without a wall outlet or external compressor.
The paper describes training-site demonstrations conducted since 2024. In one test, its 7.62-centimeter-wide body squeezed through a 3.81-centimeter opening in a steel plate. The team also demonstrated movement through broken concrete and rebar.
The public repository contains mechanical designs, circuit-board files, assembly materials, and ROS 2 software. That gives other teams a concrete starting point for confined-space robotics.
An operator controls the robot with a joystick. These demonstrations establish access through difficult terrain, not autonomous victim detection or proven disaster-response outcomes. The new paper documents the design and earlier field experience.
Read this first
Read TAO-Force. Its tool changes make a practical challenge clear: a movement that works with one object can fail when contact changes. The results show where force feedback helps and where failures remain.
Read the paper