AI & ML interests
Physical AI / Robot Deployment / Multi-modal
Recent Activity
DeepReach
Researching Data and Orchestration for Real-World Robotics
DeepReach focuses on two tightly coupled research directions:
- Manipulation-Centric Robotic Data
- DROS — Distributed Robot Operating System
Our goal is to study how robots learn and coordinate in real production environments.
🧠 Robotic Data
Egocentric Manipulation
We collect and structure multi-view, wrist-centered manipulation data for dual-arm systems.
Key properties:
- Egocentric RGB-D streams
- Action-aligned trajectories
- Skill-level segmentation
- Task-sequenced demonstrations
Designed for:
- Imitation learning
- Diffusion-based control policies
- Vision-Language-Action (VLA) models
World-Model-Based Annotation
Rather than treating perception as frame-level RGB inputs, we reconstruct structured scene representations:
- Point clouds
- Object-centric embeddings
- Spatial relations
This enables:
- Semantic task querying
- Deployment-time environment reconstruction
- Structured evaluation beyond pixel loss
We view world models as the bridge between perception and manipulation.
Manipulation as Compositional Skills
We represent tasks as compositions of atomic skills rather than monolithic policies.
This allows:
- Skill reuse across tasks
- Fine-grained failure analysis
- Scalable dataset construction
⚙️ DROS
Distributed Robot Operating System
DROS explores orchestration for heterogeneous robot fleets.
We focus on:
- Capability-aware task decomposition
- Multi-agent coordination under physical constraints
- Integration-aware scheduling across production systems
Rather than optimizing single-agent policies, we study:
How robotic capabilities compose across agents.
🔁 Closed-Loop Learning
We connect:
Deployment → Data → Model → Evaluation → Redeployment
Robots improve from real-world interaction traces rather than static benchmarks.
Research Themes
- Egocentric manipulation learning
- World-model-driven task evaluation
- Multi-agent capability graphs
- Skill composition under uncertainty
- Real-to-real adaptation in production settings
Vision
To understand how robotic systems:
- Learn from deployment
- Coordinate across heterogeneous hardware
- Transition from isolated policies to workforce-level intelligence
For collaboration and research inquiries:
contact@deepreach.ai