Introducing Perceptron Mk1

TL;DR

Perceptron Mk1 is a closed-source VLM from Perceptron aimed at video understanding and embodied reasoning, positioned to match Gemini Pro-tier quality at Gemini Flash Lite-tier price ( $0.15 per M input tokens,$ 1.50 per M output). It targets the failure modes the wiki has been tracking on frontier VLMs — pixel-precise pointing, dense counting, complex OCR (clocks, gauges, analog instruments), and structured document extraction — and treats spatial primitives (point, box, polygon, track, clip) as first-class outputs alongside text. It advertises a hand detection + 21-keypoint pose-estimation sidecar as one downstream use, and pitches Mk1 as a stack piece for both offline data curation (episode filtering, subtask boundaries, reward labels) and online use alongside a VLA (grasp affordances, cross-camera identity, video-derived success signals).

Key claims

Mk1 matches the image, video, and embodied-reasoning capabilities of frontier models including the Gemini Pro series while achieving lower cost than Gemini Flash Lite at $0.15 per million input tokens and$ 1.50 per million output [§Built for the Physical World].
Mk1 is a hybrid reasoning model with structured “thinking traces” for temporal reasoning that returns a structured breakdown of what happened and when, and lets users turn reasoning off when unneeded [§What’s New — Temporal Reasoning].
Mk1 analyzes video at a dynamic frame rate up to 2 FPS across a 32K-token context window and returns actionable timecodes for specific moments in a long stream, illustrated by auto-clipping grasp attempts in teleop episodes and flagging restocking events in warehouses in a single prompt [§What’s New — Temporal Grounding].
Mk1 supports in-context learning from a single reference image or video plus multimodal comparison over two pieces of media, without fine-tuning, labeled datasets, or a training pipeline [§What’s New — In-Context Learning].
Mk1 points more accurately than prior models on tasks including hand pose estimation, and counts reliably into the hundreds on dense scenes such as parking lots, parts on a tray, and shelf inventory, where prior models plateau or guess [§Advanced Image Reasoning — Pointing and counting].
Mk1 reads messy text, analog devices, and digital instruments accurately, including precise hour and minute hand placement on clock faces and needle positions on gauges [§Advanced Image Reasoning — Complex OCR].
Mk1 converts complex documents to HTML, JSON, or Markdown with layout, tables, and hierarchy preserved on cases including multilingual text, dense tables, and handwritten annotations [§Advanced Image Reasoning — Structured document extraction].
Mk1 is trained with structured spatial primitives — point, box, polygon, track, clip — as first-class outputs alongside text, so downstream models or policies can consume them directly [§Powering the Robotics Stack].
Mk1 holds object identity across frames through occlusion and reasons over multiple camera streams jointly, positioning it as both a data-labeling engine (subtask boundaries for hierarchical VLAs, success and failure labels for reward models, action-conditioned annotations for world model training, quality scores for episode filtering) and an inference-time companion to a VLA that returns grasp affordances, constraint checks, and relational targets [§Powering the Robotics Stack].
Six deployment archetypes are listed at launch: manufacturing/industrial inspection, media and content search/moderation, robotics onboard embodied reasoning, geospatial and critical-infrastructure analysis, security/surveillance context-aware alerting, and enhanced vision as a sidecar for text-first agents including Claude and Codex [§Real-World Applications].

Method

The post is a product announcement rather than a paper and does not disclose architecture, parameter count, training corpus, or training-compute budget. What is disclosed is the interface surface: Mk1 is a VLM whose output vocabulary is text plus structured spatial primitives (points, boxes, polygons, tracks, clips) and whose input surface is image / video / multimodal reference queries with a hybrid reasoning-on / reasoning-off toggle. Video is ingested at “up to 2 FPS” over a 32K-token context window, which sets a rough envelope of tens of seconds to a few minutes of video at usable temporal resolution.

The blog identifies two structural claims. First, Mk1 is Perceptron’s move to a closed-source model family layered on top of their existing open-source Isaac series, framed as “a step-change in image, video, and embodied reasoning” that surpasses Isaac across the board. Second, the model is being pitched as a sidecar to more than one downstream system — as annotation infra for policy training (subtask segmentation, reward-model labels, action-conditioned annotations, episode filtering), as a runtime companion to a VLA (grasp affordances, cross-view identity tracking, retry/progress signals from video), and as an enhanced-perception layer bolted onto text-first agents like Claude and Codex.

The one figure-adjacent detail with a concrete pipeline claim is the “hand detection and pose-estimation sidecar” surfaced in Wei (Jerry) Zhou’s Slack quote: per-frame bounding boxes and full 21-keypoint skeletons (wrist + four joints per finger) with left/right identity, tracked through video. This is not a Mk1 architectural feature so much as a product surface — a dedicated hand-tracking output built on top of Mk1’s spatial-primitive vocabulary.

Results

Benchmark tables are referenced in section headers (“Video Benchmark Results”, “Embodied Reasoning Benchmark Results”, “Image Benchmark Results”) but the numeric contents are not in the extracted body text at filing time — the released blog uses images/figures for the tables, which are not readable via the current fetcher. The headline framing is that Mk1 matches frontier models from Google, Alibaba, OpenAI, and Anthropic on video and embodied reasoning at significantly lower cost, and that it surpasses the open-source Isaac series across the board.

Comparators explicitly named: Gemini Pro series (matched capability), Gemini Flash Lite (undercut on price), plus generic references to Google, Alibaba, OpenAI, and Anthropic frontier VLMs. No specific per-benchmark numbers, no error bars, no ablations, no held-out generalization tests are legible from the blog’s extracted text.

Why it’s interesting

Mk1 is a closed-source, video-first, embodied-reasoning VLM aimed at the exact failure modes the wiki has been documenting under VLM Perception Failures: dense counting, pointing precision, clock/gauge reading, and structured document parsing. The three filed papers on that page — Vision Language Models are Biased, Hidden in plain sight: VLMs overlook their visual representations, Visual Structures Helps Visual Reasoning: Addressing the Binding Problem in VLMs — converge on “frontier VLMs answer from the LLM prior rather than the image” as the diagnosis; Perceptron’s positioning is the closed-source vendor’s response to that diagnosis, marketed as “counts reliably into the hundreds where prior models plateau or guess.” Without the underlying benchmark tables in a readable form, this is a positioning claim, not a validated one, but it’s the first filed product on this page that names the counting/pointing failure mode as the thing it beats.

For the Pose Estimation and Motion Capture cluster, Mk1’s advertised hand sidecar (per-frame bboxes + 21-keypoint skeletons + left/right identity, video-tracked) is a productized instance of what MoCapAnything and the FMPose3D lineage do research-side. The natural comparison partner on the wiki is HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos, which makes hand-pose-signal quality a load-bearing dependency of a manipulation policy — Mk1’s pitch is that Perceptron is now that hand-pose sidecar for anyone building on top of egocentric video.

For VLM-as-Evaluator and VLA Models, Mk1’s most concrete claim is the “annotation pipeline that used to require a team of humans runs as one prompt” framing — success/failure labels for reward models, subtask boundaries for hierarchical VLAs, action-conditioned annotations for world-model training, quality scores for episode filtering. This is the same shape as Instance Labs — Verifying Robot Learning Episode Success (VLM-as-episode-verifier) and Segmenting Robot Video into Actionable Subtasks (WGO-Bench) (VLM-as-subtask-boundary-labeler), but bundled into one closed model and one price sheet. Whether Mk1 actually clears the calibration bar those pages open — human-alignment on the verification shape, agreement with hand-annotated boundaries on the segmentation shape — is the question the blog does not answer.