fromskyfall (@ManRiraku) header
fromskyfall (@ManRiraku) profile picture
Thailand
Joined January 2022
Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)Photo by fromskyfall (@ManRiraku)
Replying to @PrismaXai
@PrismaXai put it bluntly: no robot is safe. Not the one dancing. Not the one folding laundry. You cannot sandbox a 50 kg machine or roll a bad move back. An LLM that gets it wrong gets regenerated. An arm that gets it wrong goes through the table.
Replying to @PrismaXai
@PrismaXai opened Robotics Data in limited release. 120,000+ episodes already delivered, across 1,700+ scenarios. Teleoperation and egocentric clips, reviewed to one standard, growing every month. Not a demo reel. A data bank.
1
3
14
A hallucination becomes a position error. Harmful instructions become geofences, soft stops, and the laws that already cover heavy machinery. Most of this is engineering, not philosophy. The fix starts earlier, in the data the model is allowed to copy.
1
Replying to @PrismaXai
@PrismaXai opened Robotics Data in limited release. 120,000+ episodes already delivered, across 1,700+ scenarios. Teleoperation and egocentric clips, reviewed to one standard, growing every month. Not a demo reel. A data bank.
At IROS, @PrismaXai asked one question: what makes good robotics data? Not hours. Not pretty clips. Data built for the model you are training. A VLA, a world model, and a kinematics model do not want the same set.
1
1
2
25
Teams can pull off-the-shelf sets or order collection to spec. Capacity is listed at 5,000 hours a month. The point is not more footage. It is demonstrations a VLA can actually train on. Quality scored first, then shipped.
2
At IROS, @PrismaXai asked one question: what makes good robotics data? Not hours. Not pretty clips. Data built for the model you are training. A VLA, a world model, and a kinematics model do not want the same set.
Replying to @PrismaXai
@PrismaXai is live at IROS 2026 now. Booth 1122 in Pittsburgh, through Oct 1. Tomorrow the CEO talks at 3:15 PM on one question: what makes good robotics data. That is the whole thesis on a stage.
1
1
1
23
Mix those jobs and the model learns noise. Good data matches the task, the view, and the failure cases the robot still cannot do. That is the bar PrismaX is selling at booth 1122. Quality first. Then scale.
1
4
1,096,361 1BTC = 1MBTC .-- . / .-- .. .-.. .-.. / ... . -.-. ..- .-. . / .---- --..-- ----- ----. -.... --..-- ...-- -.... .---- / ... .- - --- ... .... .. / -... - -.-. .-.-. memebitcoin.org/register
1
10
fromskyfall retweeted
Join the tide seals.garden/ (🌱)
9,392
7,644
318
8,550
304,490
Replying to @PrismaXai
@PrismaXai is live at IROS 2026 now. Booth 1122 in Pittsburgh, through Oct 1. Tomorrow the CEO talks at 3:15 PM on one question: what makes good robotics data. That is the whole thesis on a stage.
IROS 2026 starts in 3 days. Pittsburgh, Sept 28–30.@PrismaXai will be at Booth 1122. This is not a crypto meetup. It is one of the main rooms where robotics labs and builders compare what actually works.
1
1
1
46
Do not watch IROS for another robot arm. Watch whether labs agree that quality data is the bottleneck. PrismaX is in the room to argue that clean demonstrations beat raw footage. Booth first. Talk tomorrow.
1
IROS 2026 starts in 3 days. Pittsburgh, Sept 28–30.@PrismaXai will be at Booth 1122. This is not a crypto meetup. It is one of the main rooms where robotics labs and builders compare what actually works.
VLA means Vision-Language-Action. The model sees the scene, reads the task, then outputs the next move. Not a chatbot. Not a robot body. It is the brain that turns a view into work. @PrismaXai
1
1
1
73
Watch the booth for the data story, not another arm on a table. Hardware is already crowded. The scarce thing is still clean real-world training signal. That is the case PrismaX takes into IROS.
3
VLA means Vision-Language-Action. The model sees the scene, reads the task, then outputs the next move. Not a chatbot. Not a robot body. It is the brain that turns a view into work. @PrismaXai
The PrismaX loop is simple. Operators teach through real work. Validators keep bad demos out. Clean data trains the model. A better model makes the next robot run cleaner. That cycle is the product. @PrismaXai
1
1
1
42
That brain is only as good as the demos it studies. If the clips are sloppy, the action layer copies sloppy work. PrismaX is building the data path those VLA models actually train on. Sight, instruction, then motion.
3
The PrismaX loop is simple. Operators teach through real work. Validators keep bad demos out. Clean data trains the model. A better model makes the next robot run cleaner. That cycle is the product. @PrismaXai
A side camera can make a demo look perfect. The robot still acts from its own eyes. If the dataset never sees the world from the wrist or head, the model is learning the wrong view. That gap is called missing egocentric data. @PrismaXai
1
1
1
21
More machines do not close the loop. Spectator clips do not close it either. Humans teach, humans filter, then intelligence compounds across robots. Physical AI scales when this cycle keeps turning.
3
A side camera can make a demo look perfect. The robot still acts from its own eyes. If the dataset never sees the world from the wrist or head, the model is learning the wrong view. That gap is called missing egocentric data. @PrismaXai
People think teleoperation is a temporary crutch. It is not. Robots can move. They still do not know which moves are worth copying. A human in the loop is how real skill enters the dataset. @PrismaXai
1
1
1
33
Egocentric footage is the robot’s first-person record. Hands, objects, and contact as the machine actually sees them. PrismaX is pushing this next because Physical AI cannot generalize from spectator clips alone. Teach from the robot’s eyes.
5
People think teleoperation is a temporary crutch. It is not. Robots can move. They still do not know which moves are worth copying. A human in the loop is how real skill enters the dataset. @PrismaXai
Replying to @PrismaXai
@PrismaXai is heading to IROS 2026 in Pittsburgh, Sept 28–30, Booth 1122. IROS is where robotics labs and builders meet. Showing up there means the data layer wants a seat next to the hardware.
1
1
1
32
That is the point of operators. Not to babysit machines forever. To generate demonstrations the model cannot invent from a lab. Autonomy comes after the work is taught well. Remote control is the teaching step.
4
Replying to @PrismaXai
@PrismaXai is heading to IROS 2026 in Pittsburgh, Sept 28–30, Booth 1122. IROS is where robotics labs and builders meet. Showing up there means the data layer wants a seat next to the hardware.
More robots do not automatically mean better Physical AI. A hundred arms making messy demos still teach the wrong habits. One clean dataset beats a warehouse of noisy clips. The bottleneck is quality, not headcount. @PrismaXai
1
1
1
30
Physical AI is hardware-rich and data-poor. A booth at IROS is not tourism. It is PrismaX taking the quality problem into the room that ships robots. Watch what they show, not just what they post.
2
More robots do not automatically mean better Physical AI. A hundred arms making messy demos still teach the wrong habits. One clean dataset beats a warehouse of noisy clips. The bottleneck is quality, not headcount. @PrismaXai
A robot clip can look clean and still be bad training data. If the hands leave the frame, the cameras drift, or the task never finishes, the model learns the wrong lesson. Smooth is not the same as useful. @PrismaXai
1
1
1
44
That is why PrismaX treats data like the product. Operators create the work. Validators keep the junk out. Scale the standard first, then scale the fleet. Better data trains better robots. More machines alone will not.
12