# Molmoact Libero

> Run the MolmoAct LIBERO checkpoint (allenai/MolmoAct-7B-D-LIBERO-0812) as a closed-loop VLA policy for the dexterous pick-and-place segment of a task. Drives a Franka Panda in the LIBERO/robosuite OSC_POSE action space from agentview + wrist cameras, served behind a vLLM-style script speaking the openpi websocket protocol; the policy server is the bundle's own preset (no policy_id). Reads the graph-scoped observation_stream each window and terminates on a gripper open→close→open cycle, a VLM yes/no check, or max_windows. Use when a pick/place (or pick-and-drop-in-container) segment on tabletop rigid LIBERO objects is delegated to a learned policy — best steered (perceive + hover above the target) first; this is the MolmoAct alternative to pi05-libero for the same task family. NOT for deformables/cloth folding, articulated objects, non-Franka embodiments, or tasks outside the LIBERO pick-place distribution.

- Skill: `graph-robots/molmoact-libero` (Agent Skill, multi-file: 6 files)
- Install (CLI): `npx skillmds@latest add graph-robots/molmoact-libero`
- Raw SKILL.md: https://api.skillmd.com/api/skills/graph-robots/molmoact-libero/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: graph-robots (https://skillmd.com/u/graph-robots)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/graph-robots/molmoact-libero

---


# molmoact-libero

Closed-loop VLA-policy skill backed by **one model checkpoint**: AllenAI's
MolmoAct LIBERO checkpoint (`allenai/MolmoAct-7B-D-LIBERO-0812`). The skill
*is* the model — it owns its serving preset (`molmoact-libero`), so a policy
node names this skill, not a free-floating `policy_id`. The closed-loop
replan/execute/terminate body and the load-bearing LIBERO observation
encoding live in `gap.runtime.policy.run_policy_loop`; the websocket client
is resolved (and cached per preset) through the executor's `PolicyExecutor`.

This is the **MolmoAct alternative to `pi05-libero`** for the same task
family — the two are the policy A/B axis the benchmark ablates. Pick whichever
the task / experiment calls for; their capability envelope is the same.

## Capability

- **Embodiment:** Franka Panda (LIBERO/robosuite), OSC_POSE delta action
  space `[Δx, Δy, Δz, Δrx, Δry, Δrz, gripper]`. No embodiment translation
  happens in the loop — the checkpoint's native action space is forwarded to
  `sim.apply_policy_action`.
- **Tasks:** the LIBERO pick-and-place distribution — pick a tabletop rigid
  object, optionally place/drop it in a container. Works best *steered*:
  perceive the target and hover the end-effector above it (preserving the
  current rotation) before handing over, so the policy starts in-distribution.
- **Not for:** deformables / cloth folding, articulated objects, non-LIBERO
  embodiments, or tasks the checkpoint never saw. If the task is outside this
  envelope, pick a different skill or report a missing capability — do not
  delegate it here and hope.

## Serving

The bundle ships its own `server.py` and declares MolmoAct-flavored openpi
as a git dep in its own `pyproject.toml`, so the bundle is **self-contained**:
no `$GAP_OPENPI_DIR` clone, no shared venv. First-run setup is
`gap skills install molmoact-libero`, which `uv sync`s the bundle's `.venv/`
with vLLM + MolmoAct deps. The launcher then spawns the server via
`uv run --project policies/molmoact-libero -- python server.py ...` (so the
bundle's own venv activates automatically) and downloads the checkpoint from
`hf://allenai/MolmoAct-7B-D-LIBERO-0812` on first run.

The bundle's `server.py` is a **placeholder** that documents how to wire
up a vLLM-style server speaking the openpi websocket protocol; replace it
with your real serving script (e.g., from an internal MolmoAct fork) before
running the bundle for the first time. Run it yourself with
`gap policy serve molmoact-libero`. A `policies:` config entry named
`molmoact-libero` overrides the recipe (e.g. an external `url:`).

## Termination & exits

The loop exits on whichever fires first — a commanded gripper
open→close→open cycle (`gripper_cycle`, the per-item terminator for
clean-all loops; set `gripper_cycle_termination: true`), a non-empty
`termination_prompt` answered yes by the VLM (`completed_by_vlm`), or the
`max_windows` backstop. These are the subgraph's success exits; the failure
exit is `failed` (the loop raised). **Whether the task actually succeeded is
a checkpoint, not an exit** — attach a postcondition that checks the world
(e.g. the object is in the container), never an exit value like "folded".