Skip to content

MiaoDX/roboclaws

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

404 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Roboclaws

CI (main) Live Reports Python Install License

Let's Bring Brain To Robots

Visible household-robot demos driven by MCP tools, reusable skills, and AI coding agents.

Roboclaws is a thin demo repo for making AI-driven robotics behavior reviewable: frames, maps, tool traces, scores, and public/private evaluation boundaries are published as HTML reports instead of buried in terminal logs.

Surface, intent, skill, and capability profile architecture

It answers three practical questions:

  • How can an AI agent drive a robot?
  • What context and tools does the agent need?
  • What did the agent actually do in the simulated or robot-backed world?

MCP and Skill Design Principles

Roboclaws treats reusable robot behavior as skills first and MCP tools as a bounded public robot capability surface.

Principle Practice
Start from open-ended goals A user asks for work such as "clean the room" or "take useful photos"; an agent selects or creates a skill to do it.
Keep surfaces and presets separate Public commands use run::surface with named surface=..., natural-language prompt=..., and optional preset=... entries for repeated jobs.
Keep strategy in skills Skills own prompt strategy, scripts, examples, checks, and task-specific loops such as photo capture or cleanup.
Keep MCP bounded MCP tools expose semantic robot capabilities like observe, move, pick, place, and done; they should not hide a whole task behind one opaque call.
Profile public capabilities Semantic profiles describe reusable capability environments that skills can require; profiles compose by requirement, not by copying another profile's tools.
Label privileged help Simulator or demo helpers such as full object inventory and target-relative teleport are useful, but they stay labeled as privileged tools, not canonical robot abilities.
Protect private evaluation truth Hidden mess sets, acceptable destinations, private manifests, and scoring truth stay out of public profile metadata and agent-facing skill inputs.
Let reports improve skills Traces, artifacts, and evals feed the skill lifecycle: improve, split, merge, prune, or promote behavior only when the boundary is stable.

The working abstraction ladder is:

open-ended goal
  -> runnable surface and optional preset
  -> agent skill
  -> capability profile requirements
  -> MCP capability tools
  -> backend variant

Default decision: improve or add a skill when behavior changes; add or rename a surface only when the domain contract changes; add a preset= row when a known task needs its own skill, capabilities, report, or gates. Promote behavior into MCP only when multiple skills need it, the input/output shape is stable, public/private boundaries are clear, and traces can preserve the important substeps. The detailed profile and skill reference is docs/human/mcp-skills-and-semantic-profiles.md.

Run Demos With Just

Install the project once:

uv sync --extra dev

The dev extra includes the standard MolmoSpaces/MuJoCo CPU runtime used by local cleanup demos. Isaac Lab is scoped to the B1 / Map 12 digital-twin route and generic local runtime proof; keep it isolated in .venv-isaaclab/ and do not treat it as part of normal MolmoSpaces demos.

The public command grammar is named-parameter only. Public household launches name the operator-facing surface, world or scene, backend runtime, optional task preset, and agent engine separately:

just run::surface surface=<surface> agent_engine=<engine> [world=<world>] [backend=<backend>] [preset=<preset>] [prompt=<goal>] [key=value ...]

For full command routing, profiles, and maintainer-only recipes, read just/README.md.

To monitor and launch the supported local coding-agent household routes from a standalone browser console, run:

just console::run

The console uses the same world/backend/preset/agent-engine catalog for local coding-agent runs; it does not accept arbitrary browser-submitted shell commands.

Demo Matrix

GitHub Actions publishes the report site at miaodx.com/roboclaws. If a link looks stale, check the CI workflow: Pages republishes from successful main runs.

Demo What it proves Run it locally Live CI report
Semantic map build A no-cleanup sweep starts from the Base Navigation Map and builds public runtime map evidence. Online runtime_metric_map.json output and converted Agibot navigation_memory.json can both feed the canonical Actionable Semantic Map Snapshot contract. just run::surface surface=household-world world=molmospaces/val_0 backend=mujoco preset=map-build agent_engine=direct-runner evidence_lane=camera-grounded-labels camera_labeler=grounding-dino seed=7 scenario_setup=baseline Local artifact today.
Household cleanup A cleanup agent tidies a relocated household setup from Base Navigation Map context while private scoring stays hidden. just run::surface surface=household-world world=molmospaces/val_0 backend=mujoco preset=cleanup agent_engine=direct-runner evidence_lane=world-public-labels seed=7 scenario_setup=relocate-cleanup-related-objects relocation_count=5 Molmo live index, Kimi K2.6, MiMo v2.5 Pro, MiMo v2.5
Open-ended household goal A coding agent receives a user goal, builds or uses household evidence, and declares task-level completion without cleanup-specific terminal scoring. just run::surface surface=household-world world=molmospaces/val_0 backend=mujoco agent_engine=codex-cli provider_profile=codex-env prompt="find something useful to drink" Local artifact today.
Household live agent Docker-backed Claude Code or Codex connects to the cleanup MCP server and produces the same cleanup report shape. just run::surface surface=household-world world=molmospaces/val_0 backend=mujoco preset=cleanup agent_engine=claude-code provider_profile=mimo-anthropic evidence_lane=world-public-labels seed=7 scenario_setup=relocate-cleanup-related-objects relocation_count=5 Same Molmo live index; CI currently runs Claude Code through Kimi/MiMo provider profiles.
Planner proof A household cleanup run can hand off planner proof requests for local manipulation evidence without changing the public cleanup contract. just run::surface surface=planner-proof world=planner-proof/default backend=mujoco intent=planner-proof agent_engine=direct-runner mode=dry-run Local artifact today.
Agent operator console Standalone local browser console for supported Codex, Claude Code, and experimental OpenAI Agents SDK household routes with backend locks, launch-axis gates, live state, and artifact links. just console::run Local-only operator surface.
Maintainer gate Fast mock confidence check before shipping repo changes. just agent::verify mock CI status: workflow

See ARCHITECTURE.md for the code map and the full operating mode contract.

Documentation Map

Need Read
Code map and operating modes ARCHITECTURE.md
Human setup/runbooks/domain docs docs/human/README.md
Detailed MCP profile reference docs/human/mcp-skills-and-semantic-profiles.md
Skill library convention skills/README.md
Public command grammar just/README.md
Local keys and report artifacts docs/human/local-runtime.md
Coding-agent household MCP guide docs/human/coding-agent-nav-server.md
MolmoSpaces settings docs/human/molmospaces-settings.md
Current project focus STATUS.md
Agent operating rules AGENTS.md

Related Projects

  • Roboharness - visual testing harness for AI coding agents in robot simulation
  • Robowbc - whole-body-control experiments
  • OpenClaw - open-source personal AI assistant
  • ROSClaw - OpenClaw to ROS 2 bridge

License

MIT

About

Visible AI-driven robotics demos powered by VLM policies, MCP skills, and AI coding agents. Reviewable HTML reports — frames, maps, tool traces, scores — instead of buried terminal logs.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors