MOBY Robotics Deploys Edge-AI Fleets to Harvest Deep-Sea Minerals

Beneath the Surface: AI Powered Robotics for Sustainable Deep-Sea Operations With MOBY Robotics — Ai4

Ai4

Deploying vertically integrated edge-AI robotics and multi-physics simulation collapses the prohibitive capital costs of deep-sea resource extraction across Earth's unmapped biosphere.

AI-assisted summary of Beneath the Surface: AI Powered Robotics for Sustainable Deep-Sea Operations With MOBY Robotics — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published on YouTube by Ai4 on August 15, 2026, this presentation features Arjun Ramachandran, Founding Robotics Engineer at MOBY Robotics. The video description highlights how MOBY combines advanced robotics, lightweight edge-AI, and map-centric cloud intelligence to enable sustainable exploration and resource operations in extreme, unmapped deep-ocean environments.

## TL;DR Arjun Ramachandran details how MOBY Robotics addresses Earth's vast deep-ocean biosphere through a vertically integrated hardware and software stack. He introduces edge-AI platforms—specifically the 6,000-meter Scooby seabed crawler and the modular Nemo AUV—which utilize onboard NVIDIA Jetson modules for real-time perception and autonomous navigation. The talk also explores MOBY's cloud-native Ocean Intelligence Platform, which leverages 3D Gaussian Splatting and open-vocabulary models to turn dive telemetry into interactive 3D spatial maps. Finally, he covers their multi-engine simulation pipeline combining MuJoCo, Isaac Sim, and Project Chrono to train reinforcement learning policies for soft grippers and walking hexapods.

## Atmosphere & style The presentation carries a crisp, technically rigorous, and forward-looking tone suited to an advanced robotics briefing. Ramachandran speaks with the grounded pragmatism of a field engineer, fluidly connecting subsea mechanical realities like pressure vessels and terramechanics to modern computer vision and simulation workflows. The delivery feels purposeful and inspiring, framing deep-ocean robotics not as a niche enterprise, but as the essential unlocking of Earth's largest uncharted frontier.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) The deep ocean constitutes 95% to 99% of Earth's habitable biosphere volume, making ocean intelligence a fundamental planetary frontier rather than a niche market. - [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) MOBY Robotics maintains a vertically integrated US engineering and manufacturing pipeline in Miami with a 30,000-gallon test tank to satisfy domestic defense and critical mineral supply chain needs. - [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) MOBY's technical architecture is built upon three interdependent AI pillars: edge compute on autonomous vehicles, cloud-native ocean intelligence, and multi-physics synthetic simulation. - [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) The 'Scooby' seafloor crawler operates at depths up to 6,000 meters, utilizing onboard stereo vision and edge AI to autonomously harvest polymetallic nodules while protecting benthic organisms. - [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) The modular 'Nemo' AUV operates down to 300 meters, running behavioral planning and distance estimation on internal Jetson compute for infrastructure inspection and mine countermeasures. - [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) The cloud-based Ocean Intelligence Platform aggregates massive dive logs into a unified, queryable 3D archive described as an immersive 'Google Maps for the ocean.' - [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) Multi-modal AI pipelines fuse open-vocabulary detection (OWLv2), 3D Gaussian Splatting, and in-house LLMs to allow natural language search across historical dive datasets. - [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) Digital twin pipelines convert real-world dive telemetry into photorealistic synthetic scenes, enabling rapid reinforcement learning policy training while bypassing prohibitive offshore testing costs. - [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) A hybrid simulation engine stack combines MuJoCo for contact dynamics, Isaac Sim for ocean optics and sensors, and Project Chrono for seabed sediment terramechanics. - [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) Physical AI and multi-physics simulation accelerate the development of novel robotic morphologies, including bio-inspired soft octopus tentacle grippers and multi-DOF hexapod walkers.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) The Unexplored Biosphere and Vertically Integrated Marine Robotics

While standard geography curricula teach that the ocean covers roughly 70% to 71% of the Earth's surface, that figure only accounts for a flat, two-dimensional map. When measured by the total habitable volume where life can thrive, marine researchers estimate the ocean constitutes 95% to 99% of our biosphere. Every terrestrial city, forest, and piece of land-based technology ever engineered exists entirely within that remaining 1% to 5% sliver. The deep ocean represents the overwhelming majority of the living planet, yet it has remained functionally dark, inaccessible, and unmapped. Ocean intelligence is therefore not a narrow niche market, but a fundamental technical frontier.

To unlock this environment sustainably, MOBY Robotics operates out of a vertically integrated facility in Miami, Florida, maintaining an end-to-end hardware and software pipeline from initial CAD design to full ocean deployment. Designing for extreme depths requires custom hardware engineered in-house, including specialized sonars, acoustic communication modems, and pressure-balanced oil-filled (PBOF) motors. All systems undergo rigorous validation inside MOBY's on-premise 30,000-gallon test tank. Because deep-sea operations directly intersect with defense operations and the secure extraction of critical minerals, domestic manufacturing is a vital operational and compliance requirement, which simultaneously reduces unit costs across their robotic fleet.

- **Planetary Habitable Volume:** 95% to 99% of Earth's living space is subsea, leaving all terrestrial life and infrastructure within a 1% to 5% band. - **Facility & Testing:** Headquartered in Miami with an internal prototyping pipeline and a 30,000-gallon wet test tank. - **Strategic Compliance:** 100% US-based engineering and manufacturing to support domestic defense missions and supply chain requirements for critical minerals.

```mermaid flowchart TD accTitle: MOBY Robotics Three-Tier Ocean AI Architecture accDescr: Diagram showing Edge autonomy on crawlers and AUVs feeding Cloud reconstruction and informing Physics Simulation.

subgraph Edge ["1. Edge Tier: In-Situ Perception & Autonomy"] Scooby["Scooby Crawler (6,000m Depth) Stereo Vision, Nodule Harvesting, Avoidance"] Nemo["Nemo AUV (300m Depth) Side-Scan Sonar, 360° Vision, Threat ID"] JetsonCompute["Onboard Compute NVIDIA Jetson Orin Nano & NX"] Scooby --> JetsonCompute Nemo --> JetsonCompute end

subgraph Cloud ["2. Cloud Tier: Ocean Intelligence Platform"] Ingest["Dive Data Ingestion Pipeline"] VisionModels["Open-Vocabulary Models (OWLv2)"] Splatting["3D Gaussian Splatting (3DGS)"] MissionLLM["In-House LLM Mission Summaries"] JetsonCompute --> Ingest Ingest --> VisionModels Ingest --> Splatting Ingest --> MissionLLM end

subgraph Sim ["3. Simulation Tier: Synthetic Training Loop"] PhysicsEngines["Multi-Engine Simulation MuJoCo, Isaac Sim, Project Chrono"] RLTraining["RL Policy Training Soft Grippers & Hexapod Locomotion"] Cloud --> PhysicsEngines PhysicsEngines --> RLTraining RLTraining -.-> Edge end ```

[02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) Edge AI Platforms: The Scooby Crawler and Nemo AUV

Rather than treating machine learning as an afterthought bolted onto completed hardware, MOBY integrates computer vision and physical AI across three interdependent layers: edge compute, cloud intelligence, and synthetic simulation. At the edge, lightweight models operate in real time inside sealed subsea pressure vessels, enabling fully autonomous perception, path planning, and physical manipulation without human-in-the-loop tethering.

The first flagship platform is "Scooby," a heavy-duty seafloor crawler rated for depths up to 6,000 meters, engineered to navigate the seabed and harvest polymetallic nodules. Scooby utilizes forward stereo vision cameras to detect nodules, classify and avoid delicate benthic organisms, and compute real-time stereo depth estimations. These depth fields inform motion planning algorithms that coordinate an actuated front collection bucket and an onboard storage bin. Once its payload is filled, Scooby uses acoustic modems to establish relative localization within a multi-agent swarm, transiting back to a seabed smart net to deposit materials. All object detection, path planning, and obstacle avoidance models execute locally on an onboard NVIDIA Jetson Orin Nano housed in a watertight pressure enclosure.

Operating in shallower waters is "Nemo," a modular, low-cost autonomous underwater vehicle (AUV) rated for 300 meters. Nemo features a 360-degree camera rig, side-scan sonar arrays, acoustic modems, and forward stereo vision. Built to serve diverse operational profiles—ranging from marine infrastructure inspection and ocean floor mapping to defense-related threat detection and mine countermeasures (MCM)—Nemo handles behavioral state planning, local trajectory computation, and forward distance field estimation directly on internal Jetson Nano and Jetson Orin NX compute modules.

- **Scooby Specifications:** 6,000 m depth rating; automated nodule collection via actuated scoop/bin; swarm acoustic coordination; NVIDIA Jetson Orin Nano edge processing. - **Nemo Specifications:** 300 m depth rating; modular AUV architecture; side-scan sonar, stereo cameras, and 360° imaging; target inspection and mine countermeasures; NVIDIA Jetson Orin NX edge processing.

[04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) The Ocean Intelligence Platform and Cloud Reconstruction

When swarms of autonomous crawlers and AUVs surface after long missions, they offload terabytes of high-resolution sensor logs, sonar feeds, and video streams. Manual inspection of such massive data volumes is operationally unfeasible. To turn raw subsea telemetry into actionable insight, MOBY developed the Ocean Intelligence Platform—a cloud-native software suite that automatically processes, indexes, and converts incoming dive data into a unified, queryable 3D archive described as an "immersive Google Maps for the ocean."

The platform’s backend pipeline fuses multi-modal models to index the marine environment semantically and spatially. Open-vocabulary object detection models such as OWLv2 identify marine fauna, structural anomalies, and seabed assets, while 3D Gaussian Splatting (3DGS) generates photorealistic volumetric reconstructions of dive trajectories. An in-house large language model processes time-series events and vision embeddings to generate comprehensive mission reports. This architecture allows operators to execute complex natural language queries across historical fleet dives—such as asking the platform to "find dives in the coast of Florida where the octopus was spotted five times"—and immediately retrieve synchronized 360-degree footage alongside precise 3D spatial reconstructions.

- **Search & Retrieval:** Semantic querying of dive archives using open-vocabulary detection (OWLv2) linked with fine-tuned mission LLMs. - **Spatial Mapping:** Photorealistic volumetric ocean reconstructions rendered via 3D Gaussian Splatting. - **Operational Output:** Automated transformation of raw sensor feeds into interactive 3D spatial maps and mission debrief event logs.

[05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) Multi-Physics Simulation and Next-Generation Deep-Sea Manipulation

Because physical offshore field trials carry massive capital costs and operational risks, MOBY relies on high-fidelity simulation and synthetic data generation as its third core AI pillar. Real-world telemetry and environmental metadata gathered during live dives are fed back into digital twin pipelines to generate photorealistic synthetic scene variations. These synthetic environments provide the dense training ground needed to iteratively refine robotic control policies and reinforcement learning (RL) loops prior to physical deployment.

Because no single physics engine can accurately model the full complexity of marine physics, MOBY uses a hybrid simulation stack tailored to specific mechanical domains: - **MuJoCo:** Calculates precise multi-body contact dynamics and interaction forces. - **Isaac Sim (Ocean Sim):** Generates photorealistic subsea optical rendering, custom lighting dispersion, and synthetic sensor simulation. - **Project Chrono:** Models complex terramechanics and seabed sediment mobility for crawling platforms.

This synthetic simulation pipeline has enabled MOBY to design and train novel robotic morphologies that would be too costly or fragile to develop purely through physical trial-and-error. Notable examples include bio-inspired soft-body octopus tentacle grippers for non-destructive biological sampling, as well as high degree-of-freedom (multi-DOF) hexapod walking platforms built to traverse unstable, high-relief subduction zones. By combining multi-engine physics with modern physical AI, the team aims to "collapse the cost function of accessing the deep ocean," making Earth's largest living frontier fully accessible and understandable.

## Notable quotes > The deep ocean is vital to a sustainable future, yet it remains a technical abyss: no GPS, choked communications, and zero visibility. — Arjun Ramachandran

> [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) When measured by the total habitable volume where life can thrive, marine researchers estimate the ocean constitutes 95% to 99% of our biosphere. — Arjun Ramachandran

> [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) Every terrestrial city, forest, and piece of land-based technology ever engineered exists entirely within that remaining 1% to 5% sliver. — Arjun Ramachandran

> [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) The deep ocean represents the overwhelming majority of the living planet, yet it has remained functionally dark, inaccessible, and unmapped. — Arjun Ramachandran

> [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) Ocean intelligence is therefore not a narrow niche market, but a fundamental technical frontier. — Arjun Ramachandran

> [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) Rather than treating machine learning as an afterthought bolted onto completed hardware, MOBY integrates computer vision and physical AI across three interdependent layers: edge compute, cloud intelligence, and synthetic simulation. — Arjun Ramachandran

> [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) immersive Google Maps for the ocean — Arjun Ramachandran

> [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) find dives in the coast of Florida where the octopus was spotted five times — Arjun Ramachandran

> [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) collapse the cost function of accessing the deep ocean — Arjun Ramachandran

## Mentioned Arjun Ramachandran, MOBY Robotics, Ai4, Scooby, Nemo, NVIDIA Jetson Orin Nano, NVIDIA Jetson Orin NX, NVIDIA Jetson Nano, OWLv2, 3D Gaussian Splatting (3DGS), Google Maps, MuJoCo, Isaac Sim, Project Chrono, Miami, Florida

## Video description Arjun Ramachandran - Founding Robotics Engineer, MOBY Robotics

The deep ocean is vital to a sustainable future, yet it remains a technical abyss: no GPS, choked communications, and zero visibility. To operate in this extreme, high-pressure frontier, MOBY Robotics is developing autonomous systems, including seafloor crawlers, autonomous underwater vehicles (AUVs), and remotely operated vehicles (ROVs), to enable environmentally responsible access to deep-ocean resources.

This talk presents how MOBY Robotics integrates modern AI, advanced robotics, and rugged embedded systems to enable real-time perception, localization, and manipulation on the seafloor. It discusses the engineering behind robust subsea platforms, the deployment of lightweight onboard models, and the unique challenges of running edge-AI autonomy in low-visibility, high-pressure conditions, alongside the map-centric intelligence layer that turns raw dive data into actionable ocean insight.

The talk will show how advances in machine learning and robotic autonomy are shifting the paradigm and unlocking new capabilities for high-fidelity exploration, sustainable resource collection, and large-scale subsea operations.

[Source](https://www.youtube.com/watch?v=Hzo2T1ijZc4)

中文简报

MOBY Robotics 部署边缘 AI 机器人集群开采深海矿产部署垂直整合的边缘 AI 机器人技术与多物理场仿真，大幅降低了在地球尚未绘制图谱的深海生物圈中开采资源的昂贵资本成本。 Arjun Ramachandran 详细介绍了 MOBY Robotics 如何通过垂直整合的软硬件栈应对地球广袤的深海生物圈。他展示了其边缘 AI 平台——特别是下潜深度达 6,000 米的 Scooby 海底履带车以及模块化的 Nemo AUV，它们均利用板载 NVIDIA Jetson 模块实现实时感知与自主导航。演讲还探讨了 MOBY 的云原生海洋智能平台，该平台借助 3D 高斯泼溅（3D Gaussian Splatting）和开放词汇模型将下潜遥测数据转化为交互式 3D 空间地图。最后，他介绍了结合 MuJoCo、Isaac Sim 和 Project Chrono 的多引擎仿真管线，用于训练软体夹爪和六足机器人的强化学习策略。深海构成了地球宜居生物圈体积的 95% 至 99%，MOBY Robotics 在迈阿密建立了垂直整合的端到端软硬件与制造管线，并配有 30,000 加仑的测试水箱以支持国防与关键矿产供应链需求。 (0:04) MOBY 的“Scooby”海底履带车作业水深达 6,000 米，搭载 NVIDIA Jetson 边缘算力与双目视觉，实现多金属结核的自主采集并避开脆弱的底栖生物。 (2:03) 模块化自主水下航行器“Nemo”作业深度达 300 米，依靠内部 Jetson 模块运行行为规划与测距，用于海洋基础设施巡检与水雷反制。 (2:03) 云端“海洋智能平台”结合 OWLv2 开放词表检测与 3D Gaussian Splatting 技术，将海量潜水遥测数据重构成支持自然语言搜索的“海洋版谷歌地图”。 (4:24) 结合 MuJoCo、Isaac Sim 与 Project Chrono 的多物理场合成仿真管线，支持章鱼触手机械爪与多足六足机器人的强化学习训练，显著降低离岸实地测试成本。 (5:34)

Reading summary

MOBY Robotics is developing a vertically integrated hardware and artificial intelligence platform to explore and operate in the deep ocean, which accounts for an estimated 95% to 99% of Earth's habitable biosphere. By combining edge-native robotic vehicles, cloud-based volumetric mapping, and multi-physics simulation, the company aims to reduce the operational costs and risks of deep-sea resource collection and marine monitoring. Vertical Integration for Extreme Marine Environments (0:04) Although standard geographic measures state that oceans cover roughly 70% to 71% of Earth's surface, marine scientists estimate that the ocean represents 95% to 99% of the planet's habitable volume, leaving all terrestrial human infrastructure within a 1% to 5% band. Despite its scale, the deep ocean remains largely unmapped and dark. To engineer equipment capable of handling extreme pressures, MOBY Robotics maintains a vertically integrated development pipeline at its headquarters in Miami, Florida, encompassing computer-aided design, pressure-balanced oil-filled (PBOF) motors, custom sonars, acoustic modems, and validation in an on-site 30,000-gallon wet test tank. MOBY keeps its manufacturing and engineering entirely within the United States. According to the company, domestic production satisfies strict regulatory and security compliance mandates required for defense operations and the secure harvesting of seafloor critical minerals, while also lowering per-unit production costs across its fleet. Edge-Compute Robotics: Scooby Crawler and Nemo AUV (2:03) MOBY deploys edge AI directly inside sealed underwater pressure vessels to conduct real-time perception, path planning, and physical manipulation without tethered human control. The company's deep-sea flagship, "Scooby," is a seabed crawler rated for depths up to 6,000 meters. Built to gather polymetallic nodules, Scooby uses forward stereo vision cameras and an onboard NVIDIA Jetson Orin Nano to identify nodules, map stereo depth fields, and avoid sensitive benthic wildlife. An actuated collection scoop and storage bin work alongside acoustic modems, enabling Scooby to maintain relative positioning within a multi-agent swarm and deposit collected materials into a seabed smart net. For shallower water up to 300 meters, MOBY deploys "Nemo," a modular autonomous underwater vehicle (AUV). Outfitted with a 360-degree camera rig, side-scan sonar, and forward stereo vision, Nemo runs behavioral planning, distance field estimation, and trajectory tracking on NVIDIA Jetson Nano and Jetson Orin NX processors. The platform is designed for tasks including subsea infrastructure inspection, seabed mapping, and defense applications such as mine countermeasures. Cloud Reconstruction and Semantic Dive Ingestion (4:24) When subsea vehicles surface from extended operations, they offload terabytes of imaging, sonar, and environmental telemetry. To eliminate manual data review, MOBY built the Ocean Intelligence Platform, a cloud processing suite that automatically indexes raw mission data into a unified, spatial 3D archive described by the company as an "immersive Google Maps for the ocean." The pipeline applies open-vocabulary models (OWLv2) to identify marine life, structural defects, and seabed assets, while 3D Gaussian Splatting (3DGS) converts dive trajectories into photorealistic volumetric models. In-house large language models parse vision embeddings and chronological data to produce standardized mission reports. Operators can run natural language queries across historical fleet dives—such as locating instances where specific marine species were observed along a particular coastline—and retrieve corresponding 360-degree video paired with 3D spatial reconstructions. Hybrid Simulation and Novel Robotic Morphologies (5:34) Because deep-sea field trials involve severe operational risks and expenses, MOBY uses simulation to train control systems and reinforcement learning (RL) policies on synthetic data before physical deployment. To capture the full spectrum of subsea mechanics, the team employs a hybrid simulation pipeline consisting of three specialized engines: MuJoCo for multi-body contact dynamics, NVIDIA Isaac Sim for optical water effects, sensor simulation, and lighting dispersion, and Project Chrono to model terramechanics and sediment mobility on the seafloor. This simulation stack allows MOBY to evaluate experimental hardware designs without physical prototypes. Recent developments include bio-inspired soft octopus tentacle grippers designed to collect fragile biological samples without causing harm, as well as multi-degree-of-freedom hexapod crawlers designed to navigate rugged, high-relief subduction zones. By combining localized edge processing, semantic cloud mapping, and multi-engine physics simulations, MOBY Robotics seeks to substantially lower the economic barriers to deep-sea navigation and manipulation. However, validating these autonomous systems across long-duration deployments and variable benthic conditions remains an ongoing operational requirement.

精炼阅读摘要

MOBY Robotics 提出了一套融合端侧具身智能、云端三维多模态重建与多物理引擎仿真的全栈海洋机器人体系，旨在系统性降低探索占地球生物圈 95% 至 99% 可居住体积的深海环境的技术与资本门槛。深海空间认知与全垂直整合制造体系 (0:04) 按照海洋科研人员的测算，传统二维地图所展示的海洋覆盖地球表面约 70% 至 71%，但如果以生命能够繁衍的总容积（生物圈）衡量，深海占据了整个地球可居住空间的 95% 至 99%，人类陆地文明与基础设施仅存在于其余 1% 至 5% 的狭小空间内。深海虽然构成了地球生命空间的主体，却长期处于黑暗、未制图与难以触及的状态，这使海洋智能成为一项基础性的工程技术前沿。针对深海极端高压环境，MOBY Robotics 在美国佛罗里达州迈阿密建立了垂直整合的研发制造基地，覆盖从计算机辅助设计（CAD）到实际入海部署的全流程软硬件管线。为满足极限深度作业需求，团队自主研发了专用声呐、水下声学通信调制解调器以及充油压力平衡（PBOF）电机，并在其 30,000 加仑的厂区湿式测试水箱中进行全面验证。此外，由于深海勘探与国防安全及关键矿产供应链紧密相关，团队坚持 100% 的美国本土工程与制造，以满足战略合规要求并有效摊薄机器人集群的单机建造成本。边缘端无系缆作业机器人与底栖感知 (2:03) 为了摆脱对水面母船物理系缆的依赖，MOBY 将轻量化人工智能模型直接部署在深水密封耐压壳内，实现边缘端的自主感知、局部航迹规划与物理作业。其核心作业平台之一是重型海底履带机器人“Scooby”，该设备额定工作水深达 6,000 米，主要用于多金属结核的采集。Scooby 搭载前向双目立体视觉相机，可在作业过程中识别金属结核、分类并主动避开脆弱的底栖生物，同时计算实时立体深度场以协同控制前置机械铲斗与车载料斗。在装载完成后，Scooby 借助声学调制解调器实现多智能体集群相对定位，并自主导航返回海底智能网篮卸载矿料，其全部目标检测与避障算法均在内置的 NVIDIA Jetson Orin Nano 嵌入式模块上本地运行。在较浅水域，模块化低成本自主水下载具（AUV）“Nemo”的额定作业水深为 300 米。该载具配备 360 度环视相机阵列、侧扫声呐、声学调制解调器与前向双目视觉系统，专门用于海洋基础设施巡检、海床测绘、潜在水下威胁识别以及水雷对抗（MCM）等防御任务。Nemo 搭载 NVIDIA Jetson Nano 与 Jetson Orin NX 计算单元，直接在边缘端完成行为状态机决策、局部轨迹实时优化与前向距离场估算。云端海洋智能平台与三维高斯重建 (4:24) 自主水下机器人集群完成长时间潜航任务返回水面后，通常会卸载数个太字节（TB）的高分辨率传感器日志、声呐图谱与多路视频流。由于传统人工排查如此海量的数据在运营上不具备可行性，MOBY 开发了云原生的“海洋智能平台”，将原始遥测数据自动摄取、索引并转化为可交互检索的三维空间档案，被开发者比作“海洋版沉浸式 Google 地图”。该平台的后端服务融合了多模态人工智能模型：利用开放词表目标检测模型 OWLv2 自动标注海洋底栖生物、结构异常与海底人工资产；采用三维高斯溅射（3D Gaussian Splatting, 3DGS）算法，根据潜水轨迹合成逼真的空间连续体积重建；同时结合自主微调的专有大语言模型处理时间序列事件与视觉嵌入向量，自动生成潜航任务总结报告。基于这一架构，操作人员能够使用自然语言发出复杂查询指令（例如“检索在佛罗里达海岸单次潜航中发现章鱼达到五次的记录”），系统便能迅速调出时间同步的 360 度视频切片及精确的三维空间几何模型。多物理场混合仿真与新型仿生机构控制 (5:34) 真实的深海离岸试验具有极高的资本开销与环境风险，因此高保真物理仿真与合成数据生成构成了 MOBY 研发的第三大支柱。潜航收集的现实环境元数据被持续反馈回数字孪生管线，用于构建具有光度与动力学差异的合成训练场景，为强化学习（RL）算法提供密集的离线训练样本。由于单一物理引擎无法全面模拟复杂的深海环境，团队采用了多引擎协同的仿真栈：利用 MuJoCo 精确解算多体接触动力学与机械交互力；借助 Isaac Sim（Ocean Sim）模拟真实的水下光学散射特性、光照衰减与合成传感器输入；通过 Project Chrono 模拟复杂地形力学及海底松软沉积物的流变特性。这种仿真闭环大幅加速了非常规形态深海机器人的开发进度，避免了高昂的原型试错损耗。具体成果包括用于无损捕获海洋生物标本的仿生章鱼软体触手夹爪，以及具备多自由度（multi-DOF）、可在崎岖不稳定俯冲带海沟攀爬行走的六足步行机器人。开发团队认为，将多物理引擎仿真与物理 AI 相结合，能够系统性压缩进入深海的综合成本函数。 MOBY Robotics 试图通过全栈软硬件整合与合成物理仿真闭环，将深海作业从传统的低频、昂贵工程转变为可规模化部署的自主智能体系。然而，深海极端静水压下的材料持久疲劳、微弱水声通信带宽对大规模集群协同的制约，以及松软复杂底质上软体机械机构的长期工程鲁棒性，仍是该技术路径在走向完全商业化与资源开采时需要持续验证的关键问题。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 15 日在 YouTube 平台发布，演讲者为 MOBY Robotics 的创始机器人工程师 Arjun Ramachandran。视频围绕深海作业中的技术挑战展开，介绍了 MOBY Robotics 如何将物理 AI、边缘计算与先进水下机器人相结合，实现可持续的深海资源勘探与环境洞察。 ## TL;DR 演讲首先指出海洋占据了地球生物圈 95% 至 99% 的适居体积，但传统深海勘探面临着无 GPS、通信受阻与极高成本等瓶颈。MOBY Robotics 通过软硬件垂直整合，推出了 6,000 米深海采矿履带车 Scooby 与模块化自主水下航行器 Nemo，将 Jetson Orin 边缘算力封装进耐压壳体内以实现完全自主作业。在后方，团队依托 3D Gaussian Splatting、OWL-v2 及多模态 LLM 构建了支持自然语言检索的云端“海洋智能平台”，并借助 MuJoCo、Isaac Sim 和 Project Chrono 等多物理场仿真训练章鱼仿生机械爪与六足漫游机器人，最终旨在以物理 AI 彻底击穿深海准入门槛，开启负责任的深海开发新范式。 ## Atmosphere & style 整场演讲充满前沿硬核的工程极客气息与宏大深邃的探索愿景。讲者语调沉稳笃定，以极具冲击力的地球生物圈数据为切入点，层层拆解深海极端工况下的技术攻坚细节。从耐压硬件的严苛验证到前沿具身智能与多物理场仿真的无缝衔接，叙事流畅且富有前瞻性，带给观众一种亲眼见证物理 AI 正在突破地球最后未知疆界的震撼与扎实感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) 从适居体积维度衡量，海洋占地球生物圈的95%至99%，人类所有陆地文明与技术仅局限于极薄的1%至5%表层中。 - [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) MOBY Robotics 采用全流程本土化闭环基建，在自有设施（含30,000加仑室内测试水箱）中自主研发声呐与充油电机，大幅压低深海硬件研发与制造成本。 - [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) MOBY 将 AI 深度融入底层，构建了涵盖“边缘端执行、云端智能分析、仿真与合成数据生成”的三层计算机视觉与机器人架构。 - [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) 旗舰深海履带车 Scooby 耐压深度达6,000米，利用双目立体视觉与深度估计精准采集多金属结核，同时严格规避脆弱的深海生物。 - [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) 模块化自主水下航行器 Nemo 专为300米以内作业打造，集成全景相机与侧扫声呐，高性价比覆盖海底巡检、测绘与水雷反制任务。 - [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) 两款旗舰机器人均在耐压密闭舱内搭载 NVIDIA Jetson Orin 边缘算力，摆脱母船线缆依赖，独立执行水下行为规划与前向距离场构建。 - [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) 针对潜水器集群带回的TB级原始数据，云端海洋智能平台能将其快速重构为类似“深海沉浸式谷歌地图”的可交互3D数字档案。 - [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) 平台集成 3D Gaussian Splatting（3DGS）三维重建算法与 OWL-v2 开放词汇检测模型，实现海底高保真建模与零样本生物/物体识别。 - [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) 结合大语言模型自动生成结构化任务事件报告，支持科研人员通过自然语言进行精准的时空与生物出现事件检索。 - [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) 由于真实离岸深海作业成本极高，基于真实潜水数据驱动的高保真仿真与合成数据生成成为低成本策略训练的关键支柱。 - [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) 团队深度整合 MuJoCo（接触动力学）、OceanSim/Isaac Sim（逼真视效与传感器模拟）及 Project Chrono（沉积物地面力学），搭建多物理场协同仿真流。 - [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) 借助强化学习虚拟训练，MOBY 成功落地了用于无损抓取生物的章鱼仿生软体触手夹爪与适应深海破碎地形的多自由度六足机器人。 - [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) MOBY 的终极使命是通过物理 AI 与软硬件垂直整合彻底击穿深海探索的高昂成本函数，实现对深海资源与生态的可持续认知与开发。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) 重新认知海洋：从地表面积到99%的未知生物圈我们在学校念书时都学过，海洋大约覆盖了地球表面积的70%至71%。然而这仅仅是一个平面的地理投影。如果从可容纳生命活动的“实际适居空间”或体积维度来衡量，全球海洋科研人员测得的数据显示，海洋占到了整个生物圈的95%甚至高达99%。这意味着人类在陆地上所见过的一切——所有的森林、建造的所有城市、以及人类迄今工程化制造出的所有陆基技术——全部局限在地球供给的这1%到5%的极薄表层中。海洋智能绝非人工智能领域中某种小众或边缘的商业模型，它面对的是地球上绝大多数此前处于完全黑暗、人类根本无法触及的生态空间。为了打破这一深海技术绝境，总部位于迈阿密的 MOBY Robotics 正在将物理人工智能（Physical AI）推向极端前沿，其目标是在绝不破坏可持续性的前提下，让深海资源与数据不仅变得触手可及，而且具备可操作的实用价值。MOBY 将总部设在海滨，构建起一套完全打通从初始 CAD 设计、制造测试一直到实际深潜下水全流程的本土化闭环基础设施。公司在自有设施内研发声呐、声学调制解调器（Acoustic Modems）等定制传感器，设计耐高压充油电机（Pressure-Balanced Oil-Filled Motors）等执行机构，并在其容积达30,000加仑的室内试验水箱中进行极限验证。由于深海操作往往紧密关联国家防务安全与关键矿产供应链，“美国制造”（Made in USA）在这里不仅是市场宣传，更是刚性合规要求；这种端到端的垂直整合结构，也大幅压低了深海机器人及内部解决方案的研发与制造成本。 - 地球适居生物圈体积的95%至99%位于海洋，陆地文明与技术仅占1%至5%。 - MOBY Robotics 具备从 CAD 到入水试验的闭环基建，配备30,000加仑级专业测试水箱。 - 硬件自研涵盖高频声呐、水下声学调制解调器、耐高压充油电机等深海专用组件。 ### [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) 边缘实体AI落地：6000米履带车Scooby与轻型AUV Nemo MOBY 的技术架构并未将 AI 简单当作开发后期的“外挂附件”，而是深度融入到底层系统的每一个层级，重点构建了涵盖“边缘端（Edge）”、“云端（Cloud）”以及“仿真与合成数据生成（Simulation & Synthetic Data）”的三层计算机视觉与机器人智能体系。在边缘物理执行层面，其两款旗舰机器人展示了边缘端模型对严苛海洋工况的实时掌控能力。第一款旗舰产品是深海履带车“Scooby”，其最大工作下潜深度达6,000米，主要设计用于在深海平原表面爬行并采集多金属结核（Polymetallic Nodules）。Scooby 前端搭载双目立体视觉相机，实时运行目标识别模型以精准定位散落的矿石结核，同时严格规避周边脆弱的深海海洋生物；系统通过立体深度估计进行局部运动规划，控制前端的机械料斗抓取矿石并送入车体上方的电动料斗中。料斗装满后，车体会利用声学调制解调器进行水下精确定位并与蜂群编队中的其他机器人协同通信，随后返回指定的水下智能集矿网（Smart Net）倒出矿石。整套环境检测、目标避障与路径规划算法，全部打包运行在车体内密闭耐压舱中的 NVIDIA Jetson Orin Nano 嵌入式计算板上。另一款旗舰产品是自主水下航行器（AUV）“Nemo”，外观酷似一艘小型潜艇，专为水深300米以内的各类作业任务打造。Nemo 集成了360度全景相机、侧扫声呐（Side-scan Sonar）、声学调制解调器以及双目立体视觉系统，能够胜任大范围海图测绘、海底基础设施巡检、国防威胁识别以及水雷反制（Mine Countermeasures）等复杂任务。该机型的核心竞争力在于高度模块化设计与极高性价比，其板载的 Jetson Orin Nano / Orin NX 边缘算力核心全面承载了水下行为规划、局部轨迹规划以及前向距离场（Forward Distance Field）构建等全套实时运算。 ```mermaid flowchart TD accTitle: MOBY Robotics edge autonomy pipeline subgraph SensorPerception["边缘感知与输入"] CAM["双目立体视觉 / 360°全景相机"] SONAR["侧扫声呐 / 声学调制解调器"] end subgraph EdgeCompute["耐压密封计算核心 (Jetson Orin)"] DET["结核识别与生物规避模型"] PLAN["行为规划 / 局部轨迹 / 距离场"] end subgraph ActuationSwarm["控制执行与集群协作"] ACT["机械臂抓取 / 推进姿态控制"] SWARM["水下声学通信与回巢卸载"] end SensorPerception --> EdgeCompute EdgeCompute --> ActuationSwarm ``` - **Scooby 深海履带车**：耐压深度6,000米，主攻多金属结核绿色采集，集成双目立体深度估计与自适应料斗控制。 - **Nemo 自主航行器**：设计作业水深300米，主打模块化与低成本，覆盖巡检、制图与水雷反制任务。 - **边缘算力配置**：采用封装在耐压壳体内的 Jetson Orin 系列计算单元，完全摆脱对水面母船的有线算力依赖。 ### [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) 云端海洋智能平台：多源潜水数据生成可搜索3D沉浸式归档当成群的 Scooby 履带车和 Nemo AUV 在水下展开集群化作业时，单次任务就会带回数以太字节（TB）计的原始传感器数据。面对海量非结构化的声呐、点云与高分辨率视频，人工回放与审核“完全是不切实际的”，必须依靠云端自动化流水线予以结构化提炼。 MOBY 为此打造了专有的“海洋智能平台”（Ocean Intelligence Platform）。该平台由一系列高效 API、计算节点以及互相协同的多模态模型构成。一旦潜水器浮出水面，潜水数据被接入并上传云端，系统会在极短时间内将其重构为一个完全可交互、支持语义搜索的 3D 沉浸式数字档案；对客户端用户而言，其交互体验就像是在使用一款专门面向深海环境的“沉浸式谷歌地图”。云端后台整合了最前沿的开放词汇检测模型（Open-Vocabulary Detection，如 OWL-v2）、三维高斯溅射（3D Gaussian Splatting，3DGS）三维重建算法，以及自研部署的大语言模型（LLM）。系统将任务全程的所有时空轨迹与目标识别事件串联，自动编写包含丰富语义的“任务事件报告”。科研与作业人员可以直接使用自然语言提出极其具体的查询需求，例如：“找出佛罗里达海岸附近、章鱼曾出现过至少5次的所有潜水任务”。系统能够瞬间检索并调出对应潜水记录的完整360度全景影像与超高精细度的 3D 重建场景。 - 数据接入后自动通过 3D Gaussian Splatting 完成海底环境与地形的高保真三维重建。 - 利用 OWL-v2 等模型实现无预设类别的开放式物种与物体识别。 - 借助板载记录与云端 LLM 自动生成事件报告，支持复杂自然语言时空检索。 ### [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) 仿真、合成数据与具身智能：突破深海极端工况的成本枷锁离岸深海作业的单日开销极其高昂，任何在实际海况中直接调试算法的尝试都面临极大的财务与设备风险。正因如此，仿真与合成数据生成被视作技术闭环中最关键的第三个阶段。MOBY 利用真实潜水采集回来的环境数据与元数据，在虚拟世界中高精度还原出真实海底场景，并借助合成数据引擎生成光照、悬浊度、地形各异的丰富变量环境，从而在此类环境中对机器人的闭环控制策略进行低成本、高效率的强化学习训练。为了克服传统仿真器各具短板的问题，团队将多套工业级模拟软件深度整合为互补的工作流：利用 MuJoCo 求解精细的接触动力学（Contact Dynamics），利用基于 Isaac Sim 搭建的 OceanSim 实现照片级的光影渲染和水下传感器视像模拟，并引入 Project Chrono 专门模拟水下沉积物与底质沉陷的地面力学（Terramechanics）。 ```mermaid flowchart LR accTitle: Simulation and policy training loop subgraph SimTools["多物理场协同仿真套件"] MJC["MuJoCo (接触动力学)"] SIM["OceanSim / Isaac Sim (逼真渲染与传感器)"] PC["Project Chrono (深海地面力学)"] end subgraph PolicyTraining["智能算法迭代"] RL["强化学习策略训练 (RL Policies)"] CTRL["高频闭环运动控制"] end subgraph NovelHardware["新型仿生机构落地"] TENT["章鱼触手软体机械手"] HEX["多自由度深海六足漫游机器人"] end SimTools --> PolicyTraining PolicyTraining --> NovelHardware ``` 借助高保真合成环境提供的海量低成本试错空间，MOBY 成功训练并落地了在传统深海开发中极难调校的新型执行机构与运动步态。这其中包括一款用于无损抓取海洋生物与脆弱样品的类章鱼仿生软体触手夹爪，以及一款专为超深渊复杂破碎地形设计的超多自由度六足步行机器人。通过强化学习策略（RL Policies）的虚拟训练，这些机构在未经物理下水前就已掌握稳健的自适应控制能力。演讲最后回扣了开篇的核心命题：这片掌控着全球气候演变、维系着地球主要食物链、孕育着犹如科幻小说般奇异庞大生物的99%适居空间，在过去漫长的岁月中却几乎处于人类认知与勘探的盲区。MOBY Robotics 的核心使命，正是通过物理 AI、软硬件垂直整合与端到端闭环体系，“彻底击穿并大幅削减深入深海的成本函数（Collapse the cost function of accessing the deep ocean）”，让全人类能够真正负责任、可持续地理解与探索这片未知的深蓝边界。 - **多仿真引擎协同**：MuJoCo（动力学）、OceanSim / Isaac Sim（渲染/感知）、Project Chrono（土壤地面力学）。 - **新型前沿机构**：完成章鱼触手软体机械臂与多自由度六足深海行走机器人的 RL 控制策略落地。 - **终极使命**：打破深海高昂勘测成本瓶颈，以数据驱动与先进物理 AI 开启深海可持续开发新范式。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) 从可容纳生命活动的“实际适居空间”或体积维度来衡量，全球海洋科研人员测得的数据显示，海洋占到了整个生物圈的95%甚至高达99%。 — Arjun Ramachandran > [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) 海洋智能绝非人工智能领域中某种小众或边缘的商业模型，它面对的是地球上绝大多数此前处于完全黑暗、人类根本无法触及的生态空间。 — Arjun Ramachandran > [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) 在绝不破坏可持续性的前提下，让深海资源与数据不仅变得触手可及，而且具备可操作的实用价值。 — Arjun Ramachandran > [00:04](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=4s) “美国制造”（Made in USA）在这里不仅是市场宣传，更是刚性合规要求；这种端到端的垂直整合结构，也大幅压低了深海机器人及内部解决方案的研发与制造成本。 — Arjun Ramachandran > [02:03](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=123s) MOBY 的技术架构并未将 AI 简单当作开发后期的“外挂附件”，而是深度融入到底层系统的每一个层级。 — Arjun Ramachandran > [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) 面对海量非结构化的声呐、点云与高分辨率视频，人工回放与审核“完全是不切实际的”，必须依靠云端自动化流水线予以结构化提炼。 — Arjun Ramachandran > [04:24](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=264s) 对客户端用户而言，其交互体验就像是在使用一款专门面向深海环境的“沉浸式谷歌地图”。 — Arjun Ramachandran > [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) 离岸深海作业的单日开销极其高昂，任何在实际海况中直接调试算法的尝试都面临极大的财务与设备风险。 — Arjun Ramachandran > [05:34](https://www.youtube.com/watch?v=Hzo2T1ijZc4&t=334s) 彻底击穿并大幅削减深入深海的成本函数（Collapse the cost function of accessing the deep ocean），让全人类能够真正负责任、可持续地理解与探索这片未知的深蓝边界。 — Arjun Ramachandran ## Mentioned Arjun Ramachandran, MOBY Robotics, Ai4, Scooby, Nemo, NVIDIA Jetson Orin Nano, NVIDIA Jetson Orin NX, Ocean Intelligence Platform, OWL-v2, 3D Gaussian Splatting (3DGS), MuJoCo, Isaac Sim, OceanSim, Project Chrono, Smart Net ## Video description Arjun Ramachandran - Founding Robotics Engineer, MOBY Robotics The deep ocean is vital to a sustainable future, yet it remains a technical abyss: no GPS, choked communications, and zero visibility. To operate in this extreme, high-pressure frontier, MOBY Robotics is developing autonomous systems, including seafloor crawlers, autonomous underwater vehicles (AUVs), and remotely operated vehicles (ROVs), to enable environmentally responsible access to deep-ocean resources. This talk presents how MOBY Robotics integrates modern AI, advanced robotics, and rugged embedded systems to enable real-time perception, localization, and manipulation on the seafloor. It discusses the engineering behind robust subsea platforms, the deployment of lightweight onboard models, and the unique challenges of running edge-AI autonomy in low-visibility, high-pressure conditions, alongside the map-centric intelligence layer that turns raw dive data into actionable ocean insight. The talk will show how advances in machine learning and robotic autonomy are shifting the paradigm and unlocking new capabilities for high-fidelity exploration, sustainable resource collection, and large-scale subsea operations. [Source](https://www.youtube.com/watch?v=Hzo2T1ijZc4)

Source quotes

[0:04] Arjun Ramachandran: 从可容纳生命活动的“实际适居空间”或体积维度来衡量，全球海洋科研人员测得的数据显示，海洋占到了整个生物圈的95%甚至高达99%。 [0:04] Arjun Ramachandran: 海洋智能绝非人工智能领域中某种小众或边缘的商业模型，它面对的是地球上绝大多数此前处于完全黑暗、人类根本无法触及的生态空间。 [0:04] Arjun Ramachandran: 在绝不破坏可持续性的前提下，让深海资源与数据不仅变得触手可及，而且具备可操作的实用价值。 [0:04] Arjun Ramachandran: “美国制造”（Made in USA）在这里不仅是市场宣传，更是刚性合规要求；这种端到端的垂直整合结构，也大幅压低了深海机器人及内部解决方案的研发与制造成本。 [2:03] Arjun Ramachandran: MOBY 的技术架构并未将 AI 简单当作开发后期的“外挂附件”，而是深度融入到底层系统的每一个层级。 [4:24] Arjun Ramachandran: 面对海量非结构化的声呐、点云与高分辨率视频，人工回放与审核“完全是不切实际的”，必须依靠云端自动化流水线予以结构化提炼。 [4:24] Arjun Ramachandran: 对客户端用户而言，其交互体验就像是在使用一款专门面向深海环境的“沉浸式谷歌地图”。 [5:34] Arjun Ramachandran: 离岸深海作业的单日开销极其高昂，任何在实际海况中直接调试算法的尝试都面临极大的财务与设备风险。 [5:34] Arjun Ramachandran: 彻底击穿并大幅削减深入深海的成本函数（Collapse the cost function of accessing the deep ocean），让全人类能够真正负责任、可持续地理解与探索这片未知的深蓝边界。

Source quotes

Arjun Ramachandran: The deep ocean is vital to a sustainable future, yet it remains a technical abyss: no GPS, choked communications, and zero visibility. [0:04] Arjun Ramachandran: When measured by the total habitable volume where life can thrive, marine researchers estimate the ocean constitutes 95% to 99% of our biosphere. [0:04] Arjun Ramachandran: Every terrestrial city, forest, and piece of land-based technology ever engineered exists entirely within that remaining 1% to 5% sliver. [0:04] Arjun Ramachandran: The deep ocean represents the overwhelming majority of the living planet, yet it has remained functionally dark, inaccessible, and unmapped. [0:04] Arjun Ramachandran: Ocean intelligence is therefore not a narrow niche market, but a fundamental technical frontier. [2:03] Arjun Ramachandran: Rather than treating machine learning as an afterthought bolted onto completed hardware, MOBY integrates computer vision and physical AI across three interdependent layers: edge compute, cloud intelligence, and synthetic simulation. [4:24] Arjun Ramachandran: immersive Google Maps for the ocean [4:24] Arjun Ramachandran: find dives in the coast of Florida where the octopus was spotted five times [5:34] Arjun Ramachandran: collapse the cost function of accessing the deep ocean

Original transcript

[0:04] Good afternoon, everyone. Uh do you guys remember back in school that we [0:08] were taught that the ocean covers roughly 70 to 71% of the planet, right? It's true, [0:14] but that's just the surface area. It's a flat map. If you think about the actual habitable space [0:20] or the volume where life could happen, the number as measured by marine researchers across the globe [0:26] jumps to over 95 to as high as 99%. Now, think about what that means for a second. [0:32] Everything that you've ever seen on land, every forest, every city ever built, and every piece [0:36] of land-based technology that you've engineered exists within that last 1 to 5%. A very thin slice [0:42] of what our mother earth has to offer. So, when I'm going to talk about ocean intelligence today, [0:48] it isn't a niche business model for AI. It's about the vast majority of our biosphere that we call [0:54] home being functionally dark and out of reach to us until now. I'm Arjun Ramachandran, a founding [1:01] engineer at Mobius Robotics, and over the next 9 minutes, I want to show you how we at Mobius [1:06] are pushing the limits of physical AI to take this challenging and unexplored frontier [1:12] and make it not just accessible, but actionable without compromising on sustainability. [1:18] Headquartered in Miami with the ocean in our backyard, we have built an on-premise [1:23] infrastructure to close the loop from an initial CAD design all the way up to an actual dive. So, [1:29] this involves developing a lot of custom sensors like sonars, acoustic modems, uh actuators like [1:35] pressure-balanced oil-filled motors, and testing all of them within our 30,000 gallon test tank. [1:40] So, all of this happens within the facility within the US. So, when it comes to dealing with defense [1:46] and critical minerals, made in USA is not just a marketing line. It's an essential requirement. [1:52] And as an added benefit, this vertically integrated structure [1:56] allows us to reduce the cost of our robots and internal solutions and offerings. [2:03] We were able to achieve this level of maturity because we have an amazing team and a great [2:07] leadership. But also because we didn't just bolt on AI at the end. We strategically integrated it [2:12] at every layer of our stack. Uh talking beyond, you know, productivity tools and [2:17] coding assistance, we have integrated AI at three levels or three depths of computer vision AI. [2:24] At the edge, the cloud, and simulation or synthetic data generation. And I'll be going over [2:29] at each layer by giving examples from our robots and our systems. So, the cute little bot that you [2:36] see on the screen is one of our flagship robots called Scooby. Uh it's a subsea crawler rated for [2:42] 6,000 m depth. And it is intended to drive at the surface and collect polymetallic nodules. [2:49] The way this robot works is that it has a stereo vision cameras in the front which it uses [2:54] to identify nodules, to avoid marine life, and also to do stereo depth estimation so that it [3:00] can do a motion planning to use the actuated bucket in its front to capture the nodules [3:05] and put it on the actuated bin on its top. So, once the bin is full, it uses an acoustic modem [3:11] to do localization and also communicate with the other bots in the swarm and go all the way back to [3:16] the smart net and deposit the collected nodules. So, the entire set of detection and avoidance, all [3:22] the models runs within the Jetson or in Nano that sits inside a watertight enclosure within the bot. [3:31] The Nemo is the other flagship bot that we have in the autonomous underwater vehicle category [3:36] which looks like a cute little submarine. Uh it is it is equipped with a 360 camera [3:42] and a side scan sonar along with the acoustic modem and the stereo vision. [3:46] The intention of this bot is to do a wide variety of operations starting from mapping and inspection [3:51] all the way up to difference related threat identification and mine counter measures. [3:56] The main selling point of this bot is that it's the it's very modular and it's most affordable [4:03] bot in the category of underwater autonomous underwater vehicles which is deployed for 300 m. [4:10] And all of the compute, the behavioral planning, the local trajectory planning, the forward [4:16] distance field, all of that happens within the onboard Jetson Nano and NX inside the bot. [4:24] While the Scoobys and Nemos as a swarm collect terabytes of raw specs, manual review or like [4:30] analysis becomes Yeah, it's completely out of question, right? So, that's where the ocean [4:35] intelligence platform comes in. It's a framework that comprises multiple API, multiple endpoints, [4:40] and a lot of models working together in harmony so that the moment the bot comes to the surface [4:45] and a dive data is ingested, it gets converted into a complete searchable 3D archive. So, [4:51] on the client end, you can think of it something like an immersive Google Maps for the ocean. [4:59] At the back end, we have uh models like open vocabulary detection like OWL V2 [5:07] and 3D Gaussian splats for the reconstruction and in-house deployed LLM for converting the [5:13] entire mission into a event report. This event report allows you to search things like find [5:18] dives in the coast of Florida where the octopus was like spotted five times. [5:23] So, that brings you a complete 360 video and 3D reconstruction of the dives where that happened. [5:34] The third phase is, according to me, one of the most important phases because [5:40] every offshore operation is really expensive. So, the data and the metadata that we collect [5:45] with the bots is used to create photorealistic uh reconstruction of scenes and also variations [5:52] with synthetic data generation. And using those environments, we train our bot so that we can [5:58] make the control loop better. And for this, we use a suit of simulation softwares like [6:05] MuJoCo for contact dynamics, Ocean Sim uh basically Isaac Sim for photorealistic [6:10] visualization and sensors, and Project Chrono for terra mechanics. The reason why we use we use [6:15] different softwares is because none of them are like good at everything. So, we use all of them. [6:22] And as a byproduct of generating these environments, we are able to train and develop [6:28] manipulators and locomotion strategies, which otherwise would have been really difficult. [6:32] Like, for example, on the left of the screen, you see an octopus-like tentacle gripper, [6:38] which is a soft-body manipulator. Uh and on the right, you see a multi-DOF like an hexapod [6:44] walker, which is intended for subzones. Um all of this we we are able to train the RL policies [6:50] and the control policies because we generated the environments using the synthetic data. [6:58] Okay. So, let me revisit the start of this presentation when I said that we know very [7:04] little about the ocean. The same 99% of the habitable space that governs our climate, that is [7:11] responsible for a majority of our food chain, and the place where like some of the most intelligent, [7:17] most gigantic, most powerful creatures that looks like straight out of fantasy fantasy novels exist [7:24] is the same place that we know very very little about. And we, the passionate team at Mobi, [7:29] are trying to collapse the cost function of accessing the deep ocean so that we can [7:35] explore and understand it. And we would love for you to come join us and explore it with us. [7:40] With that, thank you. Uh feel free to have a chat with us after the talk. Yeah. Thank you.