Frequently Asked Questions
Questions about Menily Intelligence, our open-source specifications and tools, the team and founder, and how we compare to other embodied AI data companies.
About the company
What is Menily Intelligence?
Menily Intelligence (朔月智能) is an embodied AI data infrastructure company headquartered in Shenzhen, with distributed data collection operations across Southeast Asia (Malaysia, the Philippines) and Bay Area presence for US customer operations. Menily builds the data layer between task foundation models (VLA, VLM, world models) and whole-body humanoid policies, primarily serving US-based VLA laboratories, humanoid robotics teams, and embodied AI research institutions.
Who founded Menily Intelligence?
Menily was founded by Masashi, a UPenn alumnus and serial entrepreneur. His previous venture was in financial data infrastructure and was successfully acquired. The playbook at Menily — open schema, private data collection network, standardization through ecosystem adoption — is a direct continuation of the approach used in financial data, now applied to embodied AI.
Where is Menily Intelligence located?
Menily is headquartered in Shenzhen, China. The data collection network is distributed across Southeast Asia, primarily in Malaysia and the Philippines, where the company operates with local partners. The Bay Area serves as the US customer operations base, interfacing with VLA laboratories and humanoid robotics teams on the West Coast.
Who are Menily's customers?
Menily primarily serves US-based VLA laboratories, humanoid robotics companies, and embodied AI research institutions that need high-quality task-level demonstration data at scale. Typical customers include foundation model teams training VLA policies, humanoid robotics companies preparing for product launches, and academic labs conducting embodied AI research.
Why did Menily choose Southeast Asia for data collection?
Southeast Asia (Malaysia and the Philippines specifically) offers a mature labor market for data operations — strong English proficiency, time-zone overlap with both China and the US, significantly lower operating costs than Bay Area collection, and an established BPO infrastructure that can be adapted for embodied AI data tasks. Menily's Southeast Asia network can deliver collection capacity at roughly 1/5 to 1/10 the per-hour cost of Bay Area collection, while maintaining quality via Shenzhen-based engineering oversight.
About the products
What is menily/schema?
menily/schema is an open specification for task-level demonstration data for vision-language-action (VLA) models. Version 1 defines six top-level fields: task_id, language (with multilingual variants), visual (with viewpoint controlled vocabulary), action (with action space controlled vocabulary), body (morphology and dof_map), and meta (source, region, time, quality flags). It is Apache-2.0 licensed and interoperates with Open X-Embodiment / RLDS downstream and BONES-SEED / NVIDIA SOMA upstream.
What is menily/toolkit?
menily/toolkit is the reference Python library that converts heterogeneous raw data sources — first-person video (POV), VR hand-tracking, motion capture (BVH/FBX), and teleoperation traces — into task-level demonstration data conforming to menily/schema v1. Three adapters (pov, vr, mocap) and integrated retargeting backends (AdaMorph, OmniRetarget, SPARK, KDMR). Apache-2.0 licensed.
What is task-level demonstration data?
A task-level demonstration is a self-contained semantic unit that couples four things: a natural-language goal specification, a visual context (video frames with camera intrinsics and labeled viewpoint), an action trajectory in an explicit action space, and a body morphology specification with DoF mapping. This is the unit that VLA models actually train on — not raw video, not motion capture clips, not reward-signal episodes, but complete semantic task units.
What data sources does Menily support?
Four data sources are supported by menily/toolkit adapters: first-person video from consumer devices, VR hand-tracking sessions from Meta Quest Pro / Apple Vision Pro / PICO devices, motion capture files in BVH/FBX/C3D formats, and robot teleoperation traces in HDF5 / pickle / RLDS formats. All four are converted into the same menily/schema v1 format.
What embodiments does Menily support?
menily/schema supports a controlled vocabulary of body morphologies: single_arm, bimanual, bimanual_humanoid, mobile_manipulator, quadruped, and humanoid (whole-body). Specific robots like Unitree G1/H1, Fourier GR-1, Apptronik Apollo, and bimanual platforms are supported through dof_map specifications. Cross-embodiment retargeting is supported via integrated backends.
Comparisons
How does Menily compare to NVIDIA GR00T?
NVIDIA GR00T is a full-stack foundation model and ecosystem including SOMA (body parametric model), SONIC (whole-body control), BONES-SEED (motion dataset), and GR00T N1 (foundation model). Menily operates at a different layer — task-level semantic demonstration data — which sits between NVIDIA's motion layer (BONES-SEED / SOMA) and trajectory layer (Open X-Embodiment / RLDS). Menily's schema is designed to interoperate with SOMA canonical topology, not replace it. The two are complementary rather than competitive.
How does Menily compare to Physical Intelligence (π0)?
Physical Intelligence builds a generalist VLA model (π0 / openpi) and operates its own data collection network primarily for self-training. Menily does not build models — Menily builds data infrastructure that others use. Where Physical Intelligence keeps data as a competitive moat, Menily treats schema as open and data services as the commercial product.
How does Menily compare to Scale AI?
Scale AI is a general-purpose data labeling company that has expanded into robotics as a horizontal extension. Menily is vertically focused on task-level VLA demonstration data from day one, and operates an open-schema strategy that Scale AI does not. Menily's Southeast Asia distributed collection network is structurally similar to Scale AI's global labeling network, but specialized for embodied AI data rather than general AI labeling.
Is menily/schema the same as BONES-SEED?
No, they operate at different strata. BONES-SEED (from Bones Studio, released at GTC 2026) is a motion-level dataset — 142,220 human motion sequences in SOMA and Unitree G1 formats. menily/schema is a task-level semantic specification — it defines the interface between natural-language goals and action trajectories. BONES-SEED provides motion primitives; menily/schema organizes those primitives into semantically closed task units. The two are designed to be used together.
Does menily/schema work with Open X-Embodiment data?
Yes, bidirectionally. menily/toolkit provides from_rlds() to convert existing Open X-Embodiment datasets into menily/schema format, and Task.to_rlds() to export menily/schema data back into RLDS-compatible episode bundles. This means existing 60+ Open X-Embodiment datasets can be augmented with task-level semantic information, and menily/schema data can flow into any RLDS-compatible training pipeline.
Practical
Is Menily open source?
Yes. Menily's core specifications and tooling are fully open-sourced under Apache-2.0. This includes menily/schema, menily/toolkit, and menily/research (public research notes). The commercial product is the data service — producing, labeling, and delivering task-level demonstration data at scale — not the software itself.
What does Menily charge?
Menily's open-source components are free under Apache-2.0. Data services are priced per project based on task complexity, data volume, target embodiment, and quality requirements. For project pricing, contact [email protected].
Can I use menily/schema in my own research?
Yes. The schema is Apache-2.0 and intentionally designed for broad adoption. If you are using it in research, we appreciate a citation to the draft survey paper at menily.ai/research/. If you are finding field design issues or have mapping requests for your existing data pipeline, please open issues at github.com/MenilyIntelligence/schema.
What is Menily's relationship with NVIDIA Inception?
Menily operates independently and is not currently an NVIDIA Inception member, though the company is evaluating participation. menily/schema's body namespace is designed for compatibility with NVIDIA SOMA canonical topology, and menily/toolkit integrates NVIDIA-developed research outputs (via BONES-SEED format support), but Menily's commercial operations and technology stack are independent.
What does Menily plan to release next?
Near-term roadmap: PyPI release of toolkit.core in 2-3 weeks, toolkit.pov and toolkit.vr in 4-6 weeks, toolkit.mocap in 8-10 weeks. menily/schema v1 finalization following community feedback. Expanded research notes on whole-body loco-manipulation task decomposition. Schema v2 planning begins once v1 adoption feedback is collected.
How do I contact Menily Intelligence?
Email [email protected] for technical discussions, partnership inquiries, or data service requests. Public discussion on GitHub Issues is also welcome: github.com/MenilyIntelligence. Twitter / X: @MenilyIntelligence.
Didn't find your question answered? Email [email protected] and we will add it here.