PhyAgentOS (@phyagentos) header
PhyAgentOS (@phyagentos) profile picture

The Harness for Physical Agents. One Agent. Any Body. Build Skills Once. Then Run Anywhere. Github: PhyAgentOS/PhyAgentOS-core Website: https://phy-agent-os

Joined April 2026
Same task. Progress in action. Arrange 2、1、0、4 into the largest number and place them on the pads. First run: failed to pick up 4. Second: placed 4, struggled with 2. Final: completed 4210. Failures, retries, and success—side by side.
11
“Sort everything into three plates.” 7 objects. No categories specified. And a grasp that misses. Watch PhyAgentOS connect task reasoning, skill composition and execution feedback—from a vague instruction to a sorted table, with corrections along the way.
14
Different robots, one unified physical interface From robot arms to more physical configurations, from task execution to simulation evaluation — connecting a diverse physical world through a unified interface.
21
PhyAgent OS: Three control paradigms. One embodied system. 👇 See how three control paradigms execute the task.
1
25
⚡️ PAOS Integrates JEV: Unlocking Real-Time Capabilities for Embodied AI Harnesses ,4.3x latency reduction compared to Astra, zero performance loss 🧵🌐 phy-agent-os.x-era.com/ 🔗 github.com/PhyAgentOS/PhyAge…
1
2
97
⏳ PhyAgentOS v1.0.0 The world's first auditable, self-evolving operating system for physical agents. One OS, any body — seamless sim-to-real transfer. Composable skills. Full-scenario execution. From intelligence to the physical world. Aug 30. Stay tuned 👇 #EmbodiedAI
2
1
5
233
[RoboCasa365 Evaluation]Demonstrates the RoboCasa365 Target50 evaluation integration: launching Verification and the π0.5 policy service, issuing refrigerator-closing tasks through natural language, and recording auditable execution results.
45
[CALVIN Evaluation]Demonstrates the integration workflow for CALVIN ABC→D long-horizon task evaluation: launching the π0.5 FlowSDE policy service, creating evaluation sessions through natural language, and using the verifier to track execution and recovery results.
43
[LIBERO Evaluation]Demonstrates the session-based evaluation workflow of the OpenVLA Agent on the LIBERO benchmark: launching TargetWS and OpenPI Policy, generating task sessions, scheduling execution, and recording results.
33
[Real-World Task Chain]Humanoid/mobile robots continuously perform physical tasks in real-world rooms, including shopping, organizing, pick-and-place, bag opening, and storage, demonstrating reliable execution from planning to action.
24
[deployment]From device connection, virtual environment, configuration files, and API keys to issuing the command for Go2 to stand up and walk, this demonstrates the entire path of first-time deployment and runtime execution.
31
Hackathon @Beta_ucb recap ⚡️ Built a bunch. Learned a ton. Took home robot swag from the challenge 🏆 10/10 would do it again 🤖
61