Research
Research notes, paper explainers, and technical write-ups on robot learning and evaluation.
Bimanual Robot Manipulation: Technical Survey and Applications 2025
Survey of bimanual robot manipulation — hardware platforms, learning algorithms, data collection challenges, and industrial applications.
Showing 69 articles
Welcome to the Era of Physical Evals
In software, evals became the bottleneck — and Mercor and Fireworks are racing to build them. In the physical world, that layer doesn't exist yet. Why real-world evals (RL²) are the moat in Physical AI.
Action Chunking vs. Whole-Trajectory Imitation: A Practical Comparison
When to use action chunking (ACT) vs. whole-trajectory behavior cloning. Latency, task horizon, and training data implications.
Build vs. Buy Robot Training Data: A 2025 Decision Framework
When should your robotics team collect its own demonstration data vs. outsource to a data provider? Cost, scale, and quality tradeoffs analyzed.
Cloud Robotics Architecture Patterns for 2025
Architecture patterns for cloud-connected robot systems — edge inference, teleoperation streaming, fleet management, and data pipelines.
A Difficulty Scale for Dexterous Manipulation Tasks
Framework for classifying manipulation task difficulty — from simple pick-and-place to contact-rich assembly — with training data requirements for each level.
Foundation Models for Robot Manipulation: RT-2, OpenVLA, Octo, and π0
Survey of foundation models for robot manipulation in 2025 — architectures, training data, capabilities, and deployment considerations.
Multi-Task Robot Learning: Sharing Data Across Tasks for Better Policies
Survey of multi-task robot learning — co-training on multiple tasks, task conditioning, and when task sharing helps vs. hurts policy performance.
The Open-Source Robot Learning Ecosystem in 2025: Tools, Datasets, and Communities
Overview of the open-source robot learning landscape — frameworks, datasets, simulators, and communities driving accessible robot AI research.
Pharmaceutical Lab Automation with Robots: A Practical Guide
How pharmaceutical and biotech labs use robots for liquid handling, sample prep, high-throughput screening, and compound synthesis automation.
Physical AI Infrastructure Stack: Data Collection, Training & Deployment Layers
The five-layer infrastructure stack being built for physical AI — collection pipelines, model training, simulation, deployment, emerging standards, and the companies building each layer. Investment analysis for 2025-2026.
Real-Time Robot Inference: Edge vs. Cloud Architecture Tradeoffs
When to run robot policy inference at the edge vs. cloud — latency analysis, hardware options, model quantization, and cost comparison.
Best Robot Learning Datasets 2025
Best robot learning datasets 2025: DROID, BridgeData, Open X-Embodiment, ALOHA, LeRobot. For imitation learning, VLA fine-tuning. Robotics Center.
Building a Robot Fleet for Large-Scale Data Collection
Scaling from one robot to a fleet for parallel data collection: hardware standardization, operator training, quality control, and cost optimization.
How to Choose a Robot Arm for Your Research Lab
Practical comparison guide for selecting research robot arms: UR5e vs OpenArm vs xArm. Payload, cost, software ecosystem, and lab fit.
Contact-Rich Manipulation: Why It's Hard and How to Approach It
Understanding contact-rich manipulation: insertion, assembly, tool use. Why it challenges both classical and learned approaches, and current best practices.
TRLC-DK1 as an open-source robot learning kit
A practical guide to TRLC-DK1 as an open-source robot learning kit, including setup flow, LeRobot integration, camera paths, and bimanual expansion.
How to Collect Robot Demonstration Data
How to collect robot demonstration data for imitation learning. Teleoperation, ALOHA, data format, best practices. Robotics Center.
How to Set Up a Robot Data Collection Pipeline
Step-by-step guide to building a robust robot demonstration data collection pipeline: hardware, software, storage, quality control, and scaling.
How We Think About Real-World Evaluation
A practical framework for real-world robot evaluation across task success, repeatability, contact behavior, and operational reliability.
Human-in-the-Loop as a First-Class Learning Signal
Why human correction, intervention, and operator behavior should be treated as first-class learning signals in real robot systems.
Robot Time Fabric: Why Temporal Alignment Is the Hidden Bottleneck in Robot Learning
Temporal misalignment between cameras, joint states, and motor commands corrupts robot demonstrations as structured label noise. RCSV's hardware-synchronized data infrastructure reduces data requirements by 2-3x.
Restaurant Automation Robots: Market Reality vs. Hype in 2025
State of restaurant robotics in 2025 — which tasks are solved, which aren't, real deployment costs, and the ROI case for chain restaurants.
Robot as a Service (RaaS): Market Analysis and Procurement Guide 2025
RaaS market overview, pricing models, vendor comparison, and decision criteria for enterprises considering robot leasing vs. purchase.
Robot Data Collection Cost Breakdown: What You're Really Paying For
Full breakdown of robot training data collection costs in 2025 — hardware, operators, annotation, infrastructure, and hidden expenses.
The Robot Data Flywheel in Practice: A Real-World Case Study
How a robot data flywheel transforms policy performance — from 500 bootstrapped demos to a continuously improving pipeline, with real metrics.
Measuring Robot Demonstration Data Quality: Key Metrics and Thresholds
How to evaluate the quality of robot teleoperation demonstrations before training — success rate, smoothness, coverage, and diversity metrics.
Fine-Tuning Robot Foundation Models: A Practical Guide
How to fine-tune robot foundation models (OpenVLA, Octo, π0) on your task — data requirements, training setup, and expected performance gains.
Open Problems in Robot Learning: 10 Research Directions for 2025–2028
The most important open problems in robot manipulation learning — from sample efficiency to long-horizon planning — and what progress looks like.
Robot Manipulation Policy Evaluation: A Rigorous Methodology Guide
How to evaluate robot manipulation policies rigorously — test set design, metrics, environmental controls, and statistical significance.
The Robot Manipulation Market Landscape 2025: Companies, Investors, and Opportunities
Comprehensive overview of the robot manipulation industry — hardware vendors, software platforms, data providers, service companies, and investment landscape.
Robot Policy Failure Mode Analysis: Why Trained Robots Fail in Deployment
Systematic analysis of how trained robot policies fail — distribution shift, covariate shift, mode collapse, and recovery strategies.
Sample Efficiency in Robot Policy Learning: How Many Demos Do You Really Need?
Analysis of demonstration sample efficiency across tasks, algorithms, and robot types — with practical guidance for budgeting your data collection.
Building a Robot Teleoperation Workforce: Hiring, Training, and Management
How to recruit, train, and retain skilled robot teleoperation operators — compensation, ergonomics, quality management, and retention strategies.
How to Scale Robot Data Collection to 10,000+ Demonstrations
Engineering challenges and solutions for scaling robot teleoperation data collection from hundreds to tens of thousands of demonstrations.
Semiconductor Pick-and-Place Automation: Precision Robotics for Electronics Manufacturing
Robot systems for semiconductor and electronics manufacturing — die attach, wire bonding, PCB assembly, and inspection automation.
Sim-to-Real Transfer for Manipulation: What Works in 2025
Current state of sim-to-real transfer for robot manipulation — domain randomization, dynamics randomization, and hybrid approaches that actually work.
Teleoperation Data vs. Kinesthetic Teaching: Comparing Collection Methods
Comparison of teleoperation vs. kinesthetic (hand-guiding) robot teaching for demonstration collection — data quality, speed, task coverage, and cost.
VR Teleoperation Hardware Guide 2025: Headsets, Controllers, and Tracking Systems
Comprehensive guide to VR hardware for robot teleoperation — Meta Quest 3, HTC Vive, Valve Index, and custom exoskeleton systems compared.
Warehouse Robot Automation ROI Analysis 2025: When the Numbers Work
ROI framework for warehouse robot deployments — payback periods, headcount reduction, error rates, and capital costs for pick-and-place and fulfillment automation.
Wrist Camera vs. Fixed Camera for Robot Manipulation: Tradeoffs and Best Practices
When to use wrist-mounted cameras vs. fixed overhead/side cameras for robot manipulation — resolution, occlusion, calibration, and training data implications.
Open-Source Robot Hardware: The 2026 Landscape
Survey of open-source robot hardware platforms in 2026: OpenArm, SO-101, ALOHA, Stanford TeleOp, LeRobot. Designs, costs, and community support.
OpenArm calibration and homing guide
A practical OpenArm calibration and homing guide covering repeatable startup checks, tool changes, and what to log before policy or teleop sessions.
OpenArm MIT control gains guide
How OpenArm teams should think about MIT control gains, safety margins, iteration discipline, and tuning notes that remain usable across sessions.
OpenArm ROS 2 control guide
A practical OpenArm ROS 2 control guide covering fake hardware, real hardware bringup, controller checks, and common debugging paths.
OpenArm SocketCAN setup guide
A practical OpenArm SocketCAN setup guide covering interface checks, motor ID mapping, timeouts, and debugging patterns for builders.
OpenArm teleoperation data logging guide
How to think about OpenArm teleoperation data logging, retries, metadata, and failure replay so demonstrations stay useful for robot learning.
OpenVLA vs Octo: Which Robot Learning Model to Choose?
OpenVLA vs Octo comparison: architecture, training data, fine-tuning, deployment. Which VLA model for your robot? Robotics Center.
Orca Hand as an open-source dexterous hand platform
A practical guide to Orca Hand as an open-source dexterous hand platform, including assembly, retargeting, tactile direction, and dexterous workflow fit.
Quadruped Robot Applications: Beyond Walking
Practical applications for quadruped robots like Unitree Go2 and Boston Dynamics Spot: inspection, security, research, education, and entertainment.
How to Evaluate Robot Policies in the Real World
A practical framework for real-world robot policy evaluation: metrics, episode counts, environment variation, and statistical significance.
Robot Arm Calibration: A Practical Guide
How to calibrate robot arms for accurate manipulation: hand-eye calibration, joint offset correction, and camera-robot registration.
Robot Buying Guide for Universities and Teaching Labs
How to select and procure robots for university teaching labs and research groups. Budget planning, curriculum integration, and vendor comparison.
Robot Data Formats Compared: RLDS, LeRobot, HDF5
Comparison of robot dataset formats: RLDS (TensorFlow), LeRobot (Hugging Face), HDF5, RoboMimic. Which format to use for your data collection pipeline.
Robot Demonstration Data Quality Checklist
How to ensure your robot demonstration data is clean, useful, and compatible with modern learning pipelines.
Robot Maintenance Schedule: What to Check and When
Preventive maintenance guide for research and production robots. Daily, weekly, monthly, and annual inspection checklists.
Robot Safety Checklist Before First Deployment
Pre-deployment safety checklist for collaborative robot installations covering risk assessment, force limits, workspace zoning, and emergency stops.
Robot Teleoperation Methods Compared
Comparison of robot teleoperation interfaces: keyboard, SpaceMouse, VR controllers, leader-follower arms, gloves. Throughput, data quality, and operator fatigue.
ROS 2 vs Custom Stack: Choosing Robot Software Infrastructure
When to use ROS 2 and when to build custom robot software. Trade-offs in latency, complexity, community support, and deployment reliability.
Sim-to-Real Transfer: A Practical Guide
How to successfully transfer robot policies from simulation to the real world. Domain randomization, system identification, and fine-tuning strategies.
SO-101 as a step-by-step learning path
A practical SO-101 step-by-step learning path covering fit, buying, setup, integration, pilot design, and community-backed troubleshooting.
VLA Models Explained: What Robotics Teams Need to Know
Practical guide to Vision-Language-Action models for robot control: RT-2, OpenVLA, Octo. How they work, when to use them, and deployment considerations.
What Makes Robot Data Learning-Ready
What makes robot data learning-ready? Actionable criteria for episode structure, annotation, failure traces, and format compatibility with ACT, Diffusion Policy, and VLA models.
Why Real-World Data Beats Simulation Alone
Why real-world robot data outperforms simulation alone — distribution gaps, sim-to-real transfer failures, and how to build hybrid data pipelines that actually work.
Data Collection for Learning-Based Robotics
How we design data collection workflows for imitation learning, RL, and foundation models. Task-driven design, multimodal capture, learning-ready delivery.
OpenArm: A Data-Centric Robotic Platform for Learning-Based Manipulation
OpenArm is designed not only as a robotic manipulation platform, but as a data-native system for learning-based robotics.
PaXini PX-6AX GEN3: A Data-Native Tactile Sensing Platform
A data-native tactile sensing platform for learning-based robotics. Spatially distributed, triaxial force perception for contact understanding.
RL Environment as a Service
Real-world reinforcement learning environments for production robotics teams. Persistent, learning-ready environments backed by real hardware.
ACT vs Diffusion Policy: Which Robot Learning Algorithm Should You Use? (2025)
ACT vs Diffusion Policy: architecture deep-dive, benchmark numbers, implementation code, and a clear framework for choosing between Action Chunking with Transformers and Diffusion Policy for your robot learning task.
VLA Models Compared: RT-2, OpenVLA, Pi0, SmolVLA, RoboFlamingo (2025)
Complete 2025 comparison of Vision-Language-Action (VLA) models: RT-2, OpenVLA, Pi0, SmolVLA, RoboFlamingo. Parameters, training data, inference speed, hardware requirements, open-source status, and when to fine-tune vs train from scratch.