Force, contact and simulation | AI robotics
Three new preprints test force control for humanoids, contact-aware legged locomotion and a policy trained without real robot data.
- humanoid robotics
- contact control
- legged locomotion
- robot learning
- simulation
- AI
- news
What actually changed in the last 24 hours? Three new preprints put concrete tests around the part that often breaks once a robot leaves simulation: contact. One measures and tracks force on a humanoid, one learns legged motion across contact changes, and one claims to transfer a simulated policy to a real arm without training on real robot data. These are author-reported research results, not products or independent validation.
FLORE puts force into humanoid control
The FLORE preprint, submitted on October 26, proposes that a humanoid should do more than follow a hand trajectory. It also receives a target force for each hand and corrects it through force-torque sensing while maintaining balance.
The authors collect robot-free demonstrations with force-sensing handheld grippers and body tracking. Across three physical contact tasks, they report 80% success, compared with 26.7% for a position-only policy trained with the same demonstrations. That is a study result, not an installation-ready recipe.
The smaller-build takeaway is straightforward. If a gripper closes on something fragile, a target position is not enough. Measure a contact signal and decide how much force error is acceptable before the robot keeps pushing.
A contact simulation that reaches hardware without final tuning
Differentiable Contact Dynamics, also submitted on October 26, changes how the learning gradient is calculated around friction and contact changes. Its motion rollout keeps a hard-contact model. The smoothing happens in learning, where it gives the policy information as contact moves from support to slip.
The authors report training with 100,000 samples in under two minutes on a desktop computer, then transferring the policy to hardware without fine-tuning. On their robot, they describe trotting through pushes, debris and slippery ground. It remains a preprint submitted for review, so it is a technique to reproduce and measure, not a promise that every quadruped will behave the same way.
There is a useful rule for anyone simulating before assembly. Do not only score forward motion on clean ground. Test contact transitions, pushes and different friction. That is where a controller that looks good in a video usually sends the bill.
ZAP separates task understanding from robot motion
The third preprint, Zero-real Agent Policy, uses a pretrained vision-language model to understand a scene and then tunes it in simulation to produce actions. The authors state that this training uses no real robot data.
They report 96.7% success on their real-world tasks and transfer to an unseen Franka arm. They also compare the method with other policies inside their own benchmark. The figures are notable, but this publication has no independent verification and the preprint does not link an installable package.
The useful thing to watch is not the percentage alone. It is the split of responsibilities. A large model can help decide what needs doing. The layer that moves motors still needs small tests, physical limits and a way to measure failures. You can use that split without a lab: use vision to select an object, then simple control rules to approach, close and stop a gripper.
All three papers point at the same bottleneck. A robot leaving simulation must fail less often when it touches the world. You do not need to solve that in one go. Start with one sensor, one friction test and one force limit you can check.
- FLORE, arXiv preprint
- Differentiable Contact Dynamics, arXiv preprint
- Zero-real Agent Policy, arXiv preprint
Every claim above comes from here. If something has no source, it's not in.