Robotics Briefing /
Feeling contact, correcting movement, and walking with a different body.
Reliable robots need more than a good first guess. They need to feel resistance, correct small errors and adapt when their body changes. Three new papers test those abilities on physical hardware. In industry, faster battery replacement and more dependable network connections address two practical reasons autonomous machines stop working.
Your quick takeaways
- ForceBench shows that how a robot uses force readings matters. Two designs tied at 63% physical success, but one improved every tested task while the other hurt block insertion.
- HiRe-MoMa improved placement and sweeping after 25–30 minutes of on-robot adjustment. That time builds on prior training and still requires human supervision.
- PolyMorph used one walking policy across two physical robot platforms and several leg configurations. An unfamiliar three-legged configuration completed seven of ten outdoor trials.
Papers
ForceBench: adding force sensing is only the first step
Xuan Han and colleagues, Peking University and collaborators • October 14 preprint
A camera can show where an object sits without revealing how heavy it is or how firmly it resists movement. ForceBench compares ways to give a robot policy those missing signals. Its controlled comparison uses the same vision-language-action model, which connects images and instructions to movement, and changes how wrist force and torque enter the model.
The benchmark includes eight simulated tasks and 1,600 demonstrations. A separate physical study uses a Franka Research 3 arm, five tasks and twenty trials per task for each system. Physical policies receive their own demonstrations; this is not a direct transfer of a simulation-trained policy.
Prompt Fusion turns force readings into numerical prompt tokens alongside robot state. It completed 63 of 100 physical trials, compared with 40 for the version without force. It improved every tested task. Dual-Path Fusion, which also incorporates recent force history through another input pathway, matched the total but produced a very different result: weight sorting improved from 6 to 19 successes out of twenty, while block insertion fell from 10 to 4.
The useful lesson is that a better overall score can hide a worse contact skill. The experiments cover one physical platform with a small number of trials per task. They do not establish that one force interface will win across robots, or that simulation will always predict the physical ranking.
HiRe-MoMa: learn corrections around an existing skill
Zhefei Gong and colleagues, Northeastern, UC Riverside and Georgia Tech • October 14 arXiv release • CoRL 2026
A mobile robot must coordinate its wheels and arms. A small error in its approach can leave the gripper in the wrong place. HiRe-MoMa keeps an existing policy fixed and learns small corrections around its actions, with separate components for the mobile base and arms. The arm correction takes the base correction into account.
On a physical TIAGo++, the method raised placement success from 50% to 95% after twenty-five minutes of fine-tuning. Sweeping improved from 45% to 90% after thirty minutes. Each result comes from twenty evaluation trials. Human takeovers during training help teach the system which corrections are useful.
The harder pick-and-place task remained weak. Its reported score rose from 22.5% to 40%, but the metric awards half credit for grasping and full credit for placement. That 40% cannot be read as the proportion of complete successful runs.
The short adjustment time also sits on top of substantial preparation: the starting policies learned from eighty to two hundred demonstrations per task, and fine-tuning needs successful warm-up experience. Every training run had a supervising operator. The study uses one platform without repeated training-seed evaluations, and the authors explicitly say the policies are not ready for unsupervised use around people.
PolyMorph: one controller, several ways to walk
Hyunsik Oh and colleagues • October 14 preprint
Walking controllers usually depend on a particular robot’s joints, dimensions and mass. PolyMorph gives one shared policy a description of the robot’s body, including its links and joints. It also balances training so easier bodies do not absorb attention that harder configurations need.
The researchers tested the policy on physical Raibo1 and Raibo2 robots along the same fifty-meter outdoor route, covering flat ground, slopes, steps and uneven terrain. The four-legged group completed all twenty trials, as did the three-legged group. A two-legged Raibo2 configuration completed nine of ten; an unseen three-legged Raibo1 configuration completed seven of ten.
Success required finishing without a fall, emergency stop or operator intervention. In the three-legged hardware tests, the unused leg remained attached with locked joints, adding weight absent from the leg-removed training variants. The policy handled that difference without retraining for each configuration.
This could reduce the work needed to evaluate altered robot designs or reuse controllers. The boundary remains important: these hardware results cover two related platforms, and the authors report weaker generalization when body structure differs substantially from the training range. The unfamiliar three-legged setup failed when its swing foot caught branches or raised terrain edges.
News
DJI Dock 4 targets the time between flights
October 14 product announcement • Manufacturer measurements
DJI introduced a dock that automatically replaces a drone’s battery, reporting a twenty-five-second swap in controlled testing. Flight data transfers during the swap, and a relay feature lets two aircraft alternate coverage.
For repeated inspections or response missions, the benefit is less time waiting for a battery to recharge and less manual handling between flights. The surrounding workflow still has human oversight: a third-party alert can start a dispatch process, but an operator must confirm before the aircraft launches.
The swap figure is a manufacturer measurement, with results dependent on conditions and firmware. It does not establish continuous field availability or the time required for an entire return, landing and relaunch cycle.
FORT combines network connections for remote machines
October 14 announcement • Company-reported rollout
FORT Robotics announced software that combines cellular, satellite, Wi-Fi and radio connections into one link that adapts as network conditions change. Customers can run it on their own hardware, and FORT says rollout is underway with several defense customers.
The broader robotics problem is familiar: a remotely operated machine can lose useful visibility or control when a single network drops. Combining available connections could reduce those interruptions in worksites and other difficult environments.
The announcement gives no measured outage reduction, handover delay or independent field comparison. It describes a connectivity product and rollout, so its reliability claims still need operating evidence.
Read this first
Read ForceBench’s physical-task table. Two methods finish with the same total, yet one loses more than half the baseline’s block-insertion successes. It is a clear example of why task-level results matter when deciding what a robot can reliably do.
Read the paper