Perceptron AI is releasing Mk1.5, a real-time perception and reasoning model built to sit between raw sensor input and the control stack of physical agents like drones, quadrupeds, and smart glasses. The model accepts text, image, video, and audio input, and answers with text plus optional structured annotations such as points, boxes, polygons, and tracks , which is what lets a single checkpoint drive tracking, visual reasoning, tool use, and sub-agent dispatch without platform-specific retraining. The launch demo shows a drone given an open-ended maintenance task, exploring an unfamiliar site, keeping a gate in frame while navigating around trees, zooming in to reverse image search a logo, then locating the manual and reporting the root cause.
The context for this launch matters. Mk1.5 is positioned as Perceptron's perception model for physical agents, priced at $0.15 per million input tokens and $1.50 per million output tokens with a 36,864 token context window , which puts a capable embodied reasoning model within reach of teams that previously stitched together separate vision, planning, and tool-use systems. It follows the company's earlier Mk1 release, which focused on video understanding, temporal grounding, and embodied reasoning with structured outputs including timecodes, clips, points, and bounding boxes that are parseable at scale .
Headquartered in Bellevue, Washington, Perceptron was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former research scientists at Facebook AI Research (FAIR). The pitch to builders is straightforward. If you are shipping robotics, wearables, or autonomous hardware and want one model layer that handles perception, reasoning, and action across embodiments, Mk1.5 is the version to evaluate today.
Comments (9)
Sign in to join the discussion.
C
Chidi Nnaji18d ago
That drone-to-quadruped cut at 0:18 goes way too hard for a model release video. Whoever edited this owes us a director's cut.
M
moritz b.18d ago
The Mk1.5 wordmark is doing something weird with the '1.5' baseline. Also that gradient on the thumbnail looks like a 2019 crypto deck, please fix.
T
Takeshi M.18d ago
ok wait, one model for drones AND smart glasses AND sub-agent dispatch? either this is the most cracked demo of the quarter or the most cursed spec sheet i've read all year.
G
Greta W.18d ago
One model, many form factors sounds great in a demo reel. What does inference cost per drone-hour actually look like when the pilot ends and someone has to write a check.
E
Elin Lindqvist18d ago
Cool, but where's the API doc link? If I have to email sales to touch Mk1.5 I'm just going to keep bolting VLMs onto ROS like a caveman.
P
Priya Rao18d ago
Impressions look strong on this thread but the follow-through to the tweet after is a cliff. Feels like people bounced right when the 🧵 emoji showed up.
S
Sana Farooqi18d ago
Been watching this team quietly ship for a while. Physical AI without per-platform retraining is the actual moat, everything else is a wrapper.
D
Dan O.18d ago
'The model layer for physical AI' is fine but 'One brain. Every body.' is right there and you left it on the table.
A
Aarav Verma18d ago
Curious what the visual reasoning backbone looks like and whether the sub-agent dispatch is learned or just a tool-use scaffold on top. A tech report would be nice before we all start speculating.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Perceptron AI
The model layer for physical AI
The model layer for physical AI
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Perceptron AI is releasing Mk1.5, a real-time perception and reasoning model built to sit between raw sensor input and the control stack of physical agents like drones, quadrupeds, and smart glasses. The model accepts text, image, video, and audio input, and answers with text plus optional structured annotations such as points, boxes, polygons, and tracks , which is what lets a single checkpoint drive tracking, visual reasoning, tool use, and sub-agent dispatch without platform-specific retraining. The launch demo shows a drone given an open-ended maintenance task, exploring an unfamiliar site, keeping a gate in frame while navigating around trees, zooming in to reverse image search a logo, then locating the manual and reporting the root cause. The context for this launch matters. Mk1.5 is positioned as Perceptron's perception model for physical agents, priced at $0.15 per million input tokens and $1.50 per million output tokens with a 36,864 token context window , which puts a capable embodied reasoning model within reach of teams that previously stitched together separate vision, planning, and tool-use systems. It follows the company's earlier Mk1 release, which focused on video understanding, temporal grounding, and embodied reasoning with structured outputs including timecodes, clips, points, and bounding boxes that are parseable at scale . Headquartered in Bellevue, Washington, Perceptron was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former research scientists at Facebook AI Research (FAIR). The pitch to builders is straightforward. If you are shipping robotics, wearables, or autonomous hardware and want one model layer that handles perception, reasoning, and action across embodiments, Mk1.5 is the version to evaluate today.
Comments (9)
That drone-to-quadruped cut at 0:18 goes way too hard for a model release video. Whoever edited this owes us a director's cut.
The Mk1.5 wordmark is doing something weird with the '1.5' baseline. Also that gradient on the thumbnail looks like a 2019 crypto deck, please fix.
ok wait, one model for drones AND smart glasses AND sub-agent dispatch? either this is the most cracked demo of the quarter or the most cursed spec sheet i've read all year.
One model, many form factors sounds great in a demo reel. What does inference cost per drone-hour actually look like when the pilot ends and someone has to write a check.
Cool, but where's the API doc link? If I have to email sales to touch Mk1.5 I'm just going to keep bolting VLMs onto ROS like a caveman.
Impressions look strong on this thread but the follow-through to the tweet after is a cliff. Feels like people bounced right when the 🧵 emoji showed up.
Been watching this team quietly ship for a while. Physical AI without per-platform retraining is the actual moat, everything else is a wrapper.
'The model layer for physical AI' is fine but 'One brain. Every body.' is right there and you left it on the table.
Curious what the visual reasoning backbone looks like and whether the sub-agent dispatch is learned or just a tool-use scaffold on top. A tech report would be nice before we all start speculating.