I make machines behave.

Buraq is a reusable live agent runtime, not a one-off demo. It coordinates recognition, language generation, speech, gesture, cancellation, session state, feedback and structured interaction records while serving independent desktop, Unity and browser clients.

The architectural rule is unusually strict: clients consume the wire contract, not server code; data findings flow back to the repository that owns the cause. That discipline turns the platform into a credible research instrument.

• Working provider interfaces for ASR, LLM, TTS and multiple gesture methods.
• Streaming and queueing overlap the expensive stages instead of serializing the conversation.
• Capabilities are advertised per session; clients can be built from the tutorial and API contract alone.

OmniMo ingests motion from different file formats and skeletons and produces a consistent, provenance-aware representation for learning and runtime systems. It supports whole-corpus trees, memory-mapped reading, safe windowing, standardization, export, and build-on-demand.

The most valuable output is not only normalized data; it is a growing playbook of silent dataset failure modes discovered by distrustful measurement.

• Six processed corpora read through a common interface.
• No training window crosses a take boundary or excluded defect span.
• Quality gates exposed rigid takes, wrist flips, root defects, frame-rate errors and rest-pose ambiguities.

MotionBridge targets the practical bottleneck behind motion intelligence: turning ordinary monocular video into clean, complete, retargetable motion at a cost that a smaller lab can sustain.

• R1 learns a 3D-to-3D motion vocabulary.
• R2 aligns synthetic 2D observations with that frozen vocabulary.
• R3 distills image evidence into the token space.
• R4 benchmarks and adapts on real video with throughput, yield and usable motion-hour cost reported beside accuracy.

MotionStage is the synthetic companion to MotionBridge. It turns known skeletal motion into rendered observations while emitting the exact camera, 2D, 3D, occlusion and segmentation truth needed for controlled learning and evaluation.

It is intentionally specified around permissive tools and one consumer GPU. The first gate is one clip, end to end, with triangulation and floor-placement checks—not a glossy demo.

This review asks a question the field rarely makes answerable: what does one usable hour of motion actually cost after detection, reconstruction, cleanup, failure, licensing, compute and human labor?

The evidence base is ahead of the manuscript and continuously expanding: at least 375 unique works were citation-verified, 337 full texts were cached, and more than 343 structured notes support a reporting-practice audit. Data and figures are being completed before drafting claims.

The evaluation program spans co-speech gesture and general text-to-motion. Its central position is skeptical: a familiar metric is not automatically a meaningful instrument.

• A resource harness runs each baseline's real sampling loop and records whole-process memory and latency.
• Six of seven recent baselines are measured; the final adapter remains open.
• GQNet explores learned artifact and perceptual quality estimation; human-aligned work remains a planned study, not a published result.

ConGRets learns a global motion-derived speaker-style representation and aligns text plus style with a frozen gesture space. Runtime retrieves curated two-to-three-second units, keeping motion-capture quality while remaining friendly to chatbot and TTS pipelines.

Reported manuscript results—including 1,000 CPU queries in under 500 ms, a sub-100 MB footprint, 99% style classification and 89% content-to-cluster accuracy—remain labeled under review.

Read the archived manuscript PDF

The current design conditions co-speech motion on semantic control values such as handedness, gesture height, amplitude, speed, emotion and intensity. Prosodic signals can supervise training while text plus authored controls remain sufficient at inference.

The repository currently contains the design specification; implementation milestones and the expert study are pending. Earlier research produced a 99,022-unit control store, but that lineage is not presented as completion of the new study.

BIT proposes a standardized behavioral Turing test for embodied personas. A participant interrogates an agent that claims an identity; one behavior channel is altered, and the instrument measures both detection sensitivity and questions-to-detection.

The current repository is explicit: this is a study specification. No participants have been collected, no IRB has been filed, and runtime study instrumentation is not yet complete.

The project compares phase-aligned, fixed-rate and learned tokenizations at matched bitrate and with an identical downstream generator. The claim only survives if the phase segmenter is reliable and the tokens improve checkable semantic behavior—not merely reconstruction.

It is a staged plan. The segmenter and token analysis are explicit go/no-go gates before expensive generator training.

Instead of training a billion-parameter everything-model, this study isolates one scientific question: does text-to-action or motion-to-text data improve the semantic appropriateness of co-speech gesture at matched compute?

Five arms hold architecture, steps and token budget fixed. The study begins only after the tokenizer project freezes its representation and baseline configuration.

This project predicts phoneme durations and prosodic features from text before speech synthesis, then asks what each additional conditioning level buys in perceived synchrony, emphasis and naturalness—and what it costs in latency.

The corpus correlation phase can start independently. Larger perception work is gated by a pilot that first proves the generation backbone can express a detectable difference.

This emerging direction moves from whole-unit retrieval toward frame-level motion matching. A weighted query combines motion state, compact semantic coordinates and persona controls, preserving captured-motion realism while allowing interruption and fast redirection.

A connected skeleton-invariant proposal asks whether the same motion can be retrieved across different proportions and rigs. Both are deliberately labeled as research directions.