Agentic AI Comparison:
Code as Policies vs Unitree R1
vs
Introduction
This report compares the Unitree R1, a physical humanoid robot for research and development, with Code as Policies (CAP), a research framework for learning robot policies from code. Unitree R1 offers hardware with programmable autonomy via SDK in EDU variants, while CAP provides a software method for generating flexible robot control policies.
Overview
Unitree R1
Unitree R1 is a lightweight humanoid robot (~25kg, 1.2m tall) available in Basic ($4,900-$5,900, remote control only, no programming) and EDU variants ($10,000-$35,000 with NVIDIA Jetson Orin, full SDK, ROS2, Python/C++ support for custom development, research, and AI integration).
Code as Policies
Code as Policies (CAP) is an open-source research paradigm that trains robot control policies directly from code descriptions, enabling natural language programming of complex tasks without manual reward engineering. It runs in simulation environments like those compatible with Unitree sims.[provided URLs]
Metrics Comparison
autonomy
Code as Policies: 9
Enables high autonomy by learning end-to-end policies from code specs for complex tasks like manipulation and navigation, reducing human intervention in policy design.[provided URLs]
Unitree R1: 7
EDU variants support AI-driven mobility, autonomous navigation, voice interaction, and custom SDK programming for locomotion, SLAM, and real-time behaviors, but requires developer setup and Basic model lacks it.
CAP excels in policy-level autonomy generation; R1 provides hardware autonomy dependent on custom implementation.
ease of use
Code as Policies: 7
Uses intuitive code writing in familiar languages instead of RL hyperparameters, accessible to programmers, though needs ML training infrastructure.[provided URLs]
Unitree R1: 6
SDK supports Python/C++/ROS2 with examples, but requires expertise for setup, hardware handling, and EDU purchase; Basic is simple remote control only.
CAP is easier for software developers; R1 demands robotics/hardware skills.
flexibility
Code as Policies: 9
Highly flexible—write code for any task, generalizes across robots/environments without task-specific retraining.[provided URLs]
Unitree R1: 8
Open SDK enables deep customization (joint control, IoT, VR teleop, multi-robot), ROS2 integration, and simulation support for diverse applications.
Both highly flexible; CAP offers broader policy portability.
cost
Code as Policies: 10
Free open-source framework; only compute costs for training (no hardware purchase).[provided URLs]
Unitree R1: 7
Affordable hardware at $4,900-$35,000 (cheaper than competitors like G1), plus $500-$1,500 annual maintenance; EDU needed for advanced use.
CAP is cost-free beyond compute; R1 involves significant hardware investment.
popularity
Code as Policies: 6
Academic research method with arXiv paper and GitHub site; influential in RL/robotics communities but less widespread commercial adoption.[provided URLs]
Unitree R1: 8
Gaining traction in research/education with procurement guides, GitHub sims, YouTube demos, and developer partnerships; positioned as affordable humanoid leader.
R1 more popular as accessible hardware; CAP niche in research.
Conclusions
Unitree R1 suits hardware-centric research needing programmable physical robots (strong in flexibility/popularity, moderate cost/autonomy). Code as Policies excels for software-driven policy innovation (superior autonomy/flexibility/cost). Choose R1 for embodied experiments, CAP for scalable policy learning.[provided URLs]