ROSClaw — Give AI Agents a Body That Learns

TL;DR

ROSClaw is an open-source physical-AI runtime that sits between an agent (VLA / VLM / world model) and a real robot body, structuring every physical action through a body context → sandbox → execute → trace loop. The pitch is body-aware, guarded, replayable execution — the runtime knows which robot (e-URDF + body.yaml), validates each proposed action (ALLOW / MODIFY / BLOCK) against limits and an optional digital-twin rehearsal, executes on hardware or in sim, and persists a trace (MCAP + Parquet + JSONL) that feeds a spatiotemporal memory and a skill-promotion pipeline (Darwin). The landing page separates “ready today” (body contexts, sandbox validation, praxis trace capture), “experimental” (provider lifecycle, physical memory, hardware MCP install), and “research” (Darwin promotion, cross-body transfer, long-horizon memory) — an honest maturity split rather than a broad platform claim.

Key claims

ROSClaw is positioned as a runtime layer for embodied agents, not a policy or world model — its primitives are Body / Sandbox / Practice / Memory / Darwin / Provider, and its runtime artifacts are e-URDF + body.yaml, ALLOW / MODIFY / BLOCK decisions, MCAP + Parquet + JSONL traces, spatiotemporal evidence, and champion/candidate skill records [§04].
The system’s design bet is that every physical action becomes evidence: intent → sandbox validation → execution → trace persistence forms a single causal path, with the sandbox as the guard rail that either allows, modifies, or blocks a proposed action against body limits and a digital twin [§02].
The Hub ships three versioned registries — e-URDF Zoo (embodiment / body definitions v1.0.0, local-first), Hardware MCP Hub (agent-facing hardware adapters v0.6.0, sandbox required), and Digital Twin Hub (simulation worlds v0.4.0, local-first) — as the “installable context for real bodies” surface [§05].
Install is a single curl -sSL https://rosclaw.io/get | bash line producing ~/.local/bin/rosclaw at v0.8.2 [§03].
The homepage’s single video artifact is a Unitree G1 humanoid running a multi-step guarded action end-to-end (body: Unitree G1, guard: ALLOW + LIMITS, event: stand_balance → sandbox ALLOW_WITH_LIMITS → execute joint targets → robot state balance stable → trace persisted), framed explicitly as historical showcase rather than stock footage [§01].
Runtime maturity split: Body contexts, sandbox validation, and praxis trace capture are marked ready today; provider lifecycle (connecting VLAs / VLMs / world models / critics / classical controllers), physical memory, and hardware MCP install are experimental; Darwin skill-candidate promotion, cross-body transfer, and long-horizon memory are research [§06].
The color coding is functional rather than decorative — “physical orange” marks hardware execution, contact, warnings, and blocked actions [§04].

Method

ROSClaw is organized as a six-module runtime. Body ingests an e-URDF plus a body.yaml describing joints, sensors, tools, frames, and safety limits, so the agent knows what it is embodied as. Sandbox validates every proposed action against those limits and, when a digital twin is available, rehearses the action before permitting hardware contact; the runtime artifact is a three-valued verdict (ALLOW / MODIFY / BLOCK). Practice captures each execution as a replayable multi-format trace (MCAP for time-series signals, Parquet for tabular, JSONL for events). Memory retrieves comparable failures, recoveries, and physical context as spatiotemporal evidence to condition the next decision. Darwin compares, promotes, and rolls back skill candidates through evaluation, promoting winners to champion status. Provider is the lifecycle interface that connects heterogeneous upstream models — VLAs, VLMs, world models, critics, classical controllers — via a stable provider contract.

The Hub is the installable-context surface: e-URDF Zoo distributes body definitions, Hardware MCP Hub distributes agent-facing hardware adapters that the sandbox is required to gate, and Digital Twin Hub distributes simulation worlds, replay environments, robot assets, and regression scenes that the sandbox uses for rehearsal. All three are local-first or sandbox-required registries — an explicit choice to keep the deployment surface auditable rather than remote-executed.

Results

ROSClaw is a runtime release rather than a benchmarked model, so “results” are surface milestones. Current version is v0.8.2 (installer). The e-URDF Zoo is at v1.0.0, Hardware MCP Hub at v0.6.0, Digital Twin Hub at v0.4.0. The showcase video demonstrates a Unitree G1 humanoid completing a multi-step guarded action end-to-end under the runtime loop. No head-line benchmark numbers, throughput measurements, or robot-count adoption figures are advertised on the landing page — the page’s rhetorical move is deliberately toward evidence of execution over platform claims, with the maturity split (§06) as the honesty signal.

Why it’s interesting

ROSClaw is the second filed physical-AI runtime on the wiki (alongside DimOS — The Agentive Operating System for Physical Space (v0.0.13)), and the naming echoes the RoboClaw agentic-VLA-lifecycle stack (RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks) that the same Slack author filed the same day — suggesting a coordinated ecosystem play where RoboClaw supplies the meta-controller / EAP lifecycle recipe and ROSClaw supplies the guarded runtime that executes it on real bodies. The design-space contrast with DimOS is informative: DimOS’s primitives are Modules + Blueprints + typed pub/sub with a swappable-transport layer (LCM/DDS/ROS 2/SHM), targeting Python-first application composition by coding agents; ROSClaw’s primitives are Body + Sandbox + Practice + Memory + Darwin + Provider, foregrounding the safety-and-evidence loop (guard verdicts, spatiotemporal traces, skill-candidate promotion). Both expose an MCP-style hardware / skill surface, but DimOS treats MCP as the agent-facing API and ROSClaw treats sandbox validation as mandatory upstream of hardware. Complements Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots (Embodied.cpp is the inference runtime for VLA / WAM weights; ROSClaw is the action-execution runtime around those inferences) and pairs with VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation‘s VoLoAgent “VLA-as-interruptible-tool under a VLM orchestrator” pattern — ROSClaw’s Sandbox + Provider makes that interruption concrete at the runtime plane. Also adjacent to Inside NVIDIA Halos for Robotics: A Full-Stack Functional Safety System for Physical AI on the deployment-time safety-substrate axis: Halos is the industrial IEC-61508 SIL 3 hardware safety island, ROSClaw is the software-layer guard that would sit on top.