PhyAgentOS v1.0.0: Executable, Verifiable, Evolvable Harness for Physical Agents

PhyAgentOS v1.0.0 is an open-source harness for physical agents developed jointly by Sun Yat-sen University HCP Lab, Peng Cheng National Lab Embodied Intelligence Institute, and X-Era Lab. Released in July 2026 and stabilized in August, the project has garnered over 2,100 GitHub stars.

01 Three Robots Relay, No Full Restart on Failure

A four-minute demonstration shows a natural-language goal: "prepare and verify a six-level pH rainbow using on-site materials." The system perceives the workspace, then allocates scanning, transport, handover, and bench operations to a quadruped, a humanoid, and a dual-arm robot — all sharing a single task context. After five test tubes succeed, the sixth tube's color deviates. Instead of treating action completion as task success or restarting from scratch, the Verifier fuses live imagery and sensor readings, judges the task incomplete, appends a PlanRevision within the same AgentTask, adds buffer solution, re-tests, and only marks success when measured pH reaches 11.03 within tolerance.

02 Result Verification and Controlled Recovery: From "Action Done" to "Task Complete"

Traditional control loops equate device acknowledgment with task completion. PhyAgentOS separates execution facts, on-site evidence, and task conclusions. Query, Action, and Session record tool calls and execution states; the Evidence module stores before/after images, detection results, and device states; a task-level Verifier consumes the goal, success criteria, and evidence timeline to output success, failure, or replan-required. On verification failure, recovery stays inside the original AgentTask: the system appends a PlanRevision based on existing tool trajectories, evidence chains, and latest environment state, preserving continuity while bounding scope and retry count. The demo's single-tube rework exemplifies bounded recovery.

Benchmarks quantify the mechanism. On LIBERO's 2,000 episodes, X-VLA achieves 97.3% first-attempt success; PhyAgentOS analyzes and recovers 54 first-time failures, salvaging 26 and raising final success to 98.6% — without modifying the policy model or counting recoveries as new episodes. On CALVIN five-step long-horizon tasks, the full pipeline adds up to 4.1 percentage points. On RoboCasa365, RLDX-1 and WorldDreamer improve by 7.2 and 8.4 percentage points respectively. First-attempt and final-outcome metrics are tracked separately to distinguish raw policy capability from system-level recovery gains.

03 Experience Reflux: Verified Outcomes Drive Skill Evolution

Verification gates experience entry. Only fully verified AgentTasks become TaskEpisodes, bundling goal, plan revisions, tool trajectories, evidence fingerprints, and final verdict. Unverified or open-loop executions are excluded. Experience distillation yields two candidate types: stable, reproducible workflows become SkillCandidates; recurring, attributable issues become Lessons. Environmental anomalies (device disconnect, network jitter, sensor noise, insufficient evidence) are logged diagnostically but not added to the experience base. Skill Runtime validates artifacts, dependencies, health, and lifecycle, keeping "know-how" separate from "can-do-now."

Before activation, candidates undergo multi-task corroboration, structural validation, and safety checks. Concrete coordinates, credentials, temporary device IDs, and any instructions attempting to bypass Forge or Verifier are stripped, preventing one-off lucky successes from polluting the capability pool. The Evolution module thus accumulates traceable, condition-checkable, reusable experience — not raw logs.

04 Forge Execution Framework: Unified Physical Action Entry for Heterogeneous Robots

Forge provides a common robot-control substrate for Physical AI, abstracting robots, sensors, and environments into composable, reusable, structurally invocable capabilities. Real-world fleets mix vendors: quadrupeds for narrow-aisle inspection, arms for fixed-station pick/place, humanoids for human-designed interfaces. Previously each required custom integration. PhyAgentOS uses a unified protocol so upper-layer Agents call standardized Tools while Forge handles capability discovery, routing, command dispatch, and execution-state management.

Example: a store task "complete shelf inspection and restock in zone A" routes movement and imaging to a quadruped/wheeled robot, visual recognition to a perception module, and pick/place to an arm when conditions allow. The system tracks task completion, not manual console switching. New robot onboarding takes 5–10 minutes via capability declaration; swapping robot bodies, models, or simulators does not require rewriting task logic.

Model decision and action execution are decoupled: the Agent handles understanding, planning, and policy adjustment; Forge is the sole physical execution entry, managing capability discovery, routing, and execution state. Gateway registers and manages skills; ToolEndpoint maps tool calls to concrete robot actions; Dora dataflow carries real-time data and control. Query reads state synchronously; Action and Session trigger physical execution, distinguishing "observe world" from "change world" at the protocol level.

v1.0.0 redefines the state protocol: AGENTS.md, SKILL.md remain human/Agent-readable knowledge; real actions execute exclusively via Forge Tool API; AgentTask, PlanRevision, Evidence, and verification results are written as structured fact records. Knowledge, execution, and facts each reside in appropriate media — no more file-polling to drive physical actions. Even on timeout, cancellation, or unknown result, the system first reconciles live state before deciding next steps, reducing duplicate execution risk.

05 Composing Tools and Skills for Complex Long-Horizon Tasks

Real work decomposes into verifiable Tools: "read state," "move to shelf," "capture image," "grasp item," "place item," "verify result." Repeatedly validated sequences crystallize into Skills. This composition lets the system adapt to "inspect zone B shelves and handle out-of-stock items" by dynamically chaining move, scan, identify, grasp, verify — adjusting only the problematic step when occlusion, misplacement, or grasp failure occurs, without restarting the entire flow. Verified capability modules enable practical long-horizon execution: robots need not memorize full scripts but invoke proven modules under task constraints and push forward with live feedback.

06 Ecosystem and Integration: Keeping Models, Skills, and Robots Composable

The unified Tool contract lowers cross-configuration adaptation cost. Onboarding a new robot morphology typically takes 5–10 minutes; swapping models, simulators, or robot bodies leaves upper-layer Agents and task chains untouched. Officially supported morphologies grew from 19 to 43, covering arms, quadrupeds, humanoids, mobile manipulators, and dexterous hands, with 9 running on real hardware. The release provides 9 algorithm nodes spanning VLA, world models, visual perception, and imitation learning, with 11 more planned.

PhyAgentOS is MIT-licensed; the core repo moved to the PhyAgentOS organization as PhyAgentOS-core. After configuring a model provider, a generic Agent runs without real hardware; driving physical robots requires installing corresponding Forge Skills, Runtime, and hardware dependencies. Minimal install and launch:

git clone https://github.com/PhyAgentOS/PhyAgentOS-core.git
cd PhyAgentOS-core
python -m pip install -e .
paos onboard
paos agent

PhyAgentOS v1.0.0's core contribution is not another model wrapper but a harness that makes physical agents executable, verifiable, recoverable, and evolvable. Agent, Forge, Evidence, Verifier, and Evolution own planning, execution, evidence capture, acceptance, and experience distillation respectively — a responsibility boundary that determines whether a demo becomes a long-term, stable system capability.

Industrialization won't happen overnight; robot reliability, site safety, deployment cost, data quality, network conditions, and legacy system integration still gate embodied AI's entry into production. Yet the industry direction is clear: competition shifts from "can a robot pull off a flashy demo" to "can it stably complete real tasks, verify outcomes, handle exceptions, and continuously reduce human intervention."

In retail, after-hours shelf inspection is a typical scenario: a robot traverses aisles, reads shelf state, detects gaps or misplacements, then calls appropriate devices for verification or restock — starting with one zone, one SKU class, one time window. In factories, line changeover fits the same collaboration: mobile robots deliver materials, arms handle loading/assembly, vision checks results; on detecting a misoriented part, positional offset, or incomplete assembly, the system preserves state and reworks only the faulty step instead of halting the whole line.

X-Era Lab (拓元智慧), a primary PhyAgentOS developer, is extending this framework into smart retail, industrial manufacturing, and domestic services, letting Physical Agents continuously verify, reflux experience, and iterate capabilities — evolving from isolated demos into long-term deployable, continuously improving, replicable system capabilities. PhyAgentOS serves researchers (quantifiable embodied AI testbed), robot OEMs (full-spectrum adaptation, rapid deployment), and end users (continuous learning, stable efficient operation in complex dynamic scenes).