Today's most important lesson: better data beats more training steps.
By improving object tracking, validating lift and grip hold, and automatically filtering out failed demos, I transformed a noisy dataset into a small set of training-ready episodes with an 88.5% success rate.
Now I train with confidence, not guesswork. 🤖
#robotics #ai
What changes when a robot can actually feel an object? 🤖
Throughout my robotic grasp-learning experiments at HumxAI, I’ve seen a clear evolution in the information available to train policies:
1️⃣ Vision only
The model can observe the hand and object, but must infer when contact happens. A grasp may look correct even when the fingers are not touching the object or holding it securely.
2️⃣ Vision and motion
Adding hand trajectories helps the model understand how to approach an object. However, the exact moment of contact remains ambiguous.
3️⃣ Vision, motion, and contact sensors
Contact sensors provide direct signals about critical events:
- When contact begins
- Which fingers are involved
- Whether the object is being held or merely touched
- When the object begins to slip
- How much pressure is needed for a stable grasp
This helps me distinguish between the key stages of a demonstration: approach, contact, closure, manipulation, and release.
My main conclusion so far is that contact sensors do not replace vision; they complement it with a less ambiguous signal about what is physically happening.
I’ve also learned that adding more sensors does not automatically produce a better model. The signals must be synchronized correctly, noise must be filtered, and the dataset’s actions must remain physically coherent.
My next step is measuring how much contact information improves grasp generalization across objects, positions, and materials the model has never seen before.
I’m working toward robots that don’t just see how we perform a task, but can also feel when they are doing it correctly.
#Robotics #robotlearning #embodiedai #machinelearning #artificialintelligence #HumxAI
A lot of human-to-robot datasets struggle because video alone often misses the physical signals behind the task.
The hard part is not only seeing the human motion, but capturing what makes it transferable: contact, timing, pressure, wrist movement, and object interaction.
Without those signals, many demos look useful but are hard to turn into reliable robot behavior. #Robotics #RobotLearning
Robot learning doesn’t just need more data.
It needs data that can transfer.
A recent paper, “OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learning,” makes this point very clear:
Generalist robot policies need to work across different robot arms, grippers, tasks, and environments.
But most robot datasets are still biased toward a small number of embodiments.
That matters.
A policy can look strong in one setup, but fail when the arm changes, the gripper changes, the object changes, or the environment becomes less controlled.
This is why cross-embodiment generalization is becoming one of the most important problems in robot learning.
The goal is not just:
“Can this robot do the task?”
The real question is:
“Can this skill transfer to another robot, another gripper, another object, and another real-world condition?”
Open datasets like OXE and OXE-AugE are pushing the field forward.
But every robotics company still has its own long tail of failures:
polybags, cables, deformable objects, odd grasps, kitting, stacking, handoffs, tools, fixtures, transparent objects, reflective parts, and customer-specific edge cases.
That long tail is hard to solve with generic data alone.
This is the problem we’re working on at HumxAI.
We collect targeted human manipulation demos and convert them into robot-learning datasets that can be retargeted to different robot end-effectors — from simple grippers to humanoid hands.
Our focus is not generic video.
It’s high-signal manipulation data:
RGB-D/LiDAR, per-finger contact, wrist IMU, object tracking, and task-specific failure modes.
The way I see it:
Large open datasets help robots learn broad capabilities.
Targeted datasets help robots overcome the specific failures that block real deployment.
HumxAI is building the data layer for that long tail.
La #InteligenciaArtificial simplemente NO puede reemplazar los copywriters, si todos los copys se crear con IA, dejarían de tener efecto... Por cierto, la imagen fue creada con IA.🤣 #newsletter#copywriting
Validating a business idea? Don't forget to check the search traffic! It's a valuable way to gauge interest and demand. One option is https://t.co/m6DdcjTjeX #BusinessTips#Entrepreneurship
3️⃣ Build a minimum viable product (MVP): Create a prototype or simplified version of your product/service to test its viability and gather feedback from potential clients.
4️⃣ Seek feedback from experts: Connect with industry experts, mentors, or advisors who can provide valuable insights and help refine your business idea.
5️⃣ Engage with your target audience: Conduct surveys, interviews, or focus groups with your potential B2B customers to validate your assumptions and gather real-world feedback.
6️⃣ Pilot test with early adopters: Offer your product/service to a select group of customers willing to try it out, gather their feedback, and iterate accordingly.
ChatGPT and AI threads are overused.
But there's 1000s of genuinely valuable AI tools out there.
Here's 8 AI websites I actually use to make my life easier: