MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation

What it claims

MobiAgent pairs a VLM planner–executor–critic loop at deployment with an offline loop that turns its own rollouts into training data for the skills. On BEHAVIOR-1K it reaches a 65.0% mean success rate, 22.5 points above π0.5-TA. The paper identifies π0.5-TA as the task-specific 2025 BEHAVIOR Challenge winner, which uses time-indexed stages that share a single prediction head. The authors argue that this shared head causes the capacity interference MobiAgent is built to avoid. The paper is a CoRL 2026 paper, and code is listed at https://github.com/kaiknower/MobiAgent.

How it works

Executor. The executor is built on π0.5. Every skill shares one VLM trunk, and each atomic skill (e.g. pick, place, move_to, open) gets its own flow-matching action expert: a_{t:t+L} = π_action^(s_k)(π_vlm(ℓ_k, o_t)). Sub-tasks that use the same underlying action reuse the same expert.

Planner. The planner is a VLM that runs in a receding-horizon fashion. It is called again after each skill finishes or fails. Its inputs are:

• the global instruction
• the current multi-camera observation
• an episode memory holding past sub-tasks, critic verdicts and cues such as spatial misalignment
• an open-vocabulary skill catalog

It outputs a short instruction (e.g. "pick the red can") plus a skill hint, and the hint selects the expert.

Critic. After every C action chunks, the critic looks at three frames sampled with a stride: o_{t−2CL}, o_{t−CL} and o_t. It returns a verdict, a recommendation and audit cues. The output must follow a fixed schema, and any violation is coerced to error. A deterministic orchestrator then maps the critic's output to an action:

flowchart TD
  P[Planner: next sub-task + skill hint] --> E[Skill expert runs C chunks]
  E --> C{Critic verdict}
  C -->|complete| A[ADVANCE] --> P
  C -->|error or replan| R[REPLAN] --> P
  C -->|incomplete + retry| N{n_t < n_max?}
  N -->|yes| T[RETRY] --> E
  N -->|no| R

Outer loop. This loop runs offline:

1. A VLM Data Curator segments startup demonstrations and deployment rollouts into clips, then verifies and captions each clip. Failed clips are dropped.
2. A Skill Generator clusters the captions into a discrete skill vocabulary. For example, "Grasp storage container" and "Grab container from floor" both map to pick.
3. A Skill Trainer fine-tunes the per-skill experts on the resulting data.

No human annotation is involved.

Evidence

• BEHAVIOR-1K: 65.0% mean success against π0.5-TA, a gap of +22.5 pp. The paper says component ablations show that dynamic planning, skill-specific experts and visual verification each matter, but the fetched text contains no ablation numbers.
• Outer-loop recycling: success goes from 7.50% to 27.50% on RoboCasa and from 32.5% to 57.5% on the Astribot S1 robot.
• Real-world tasks: four tasks on Astribot S1, namely paper, bottle and can disposal, and pouring blue particles into a cup.
• Error recovery: shown qualitatively, for example re-approaching and re-grasping a can that was dropped outside the trash bin.

How to use it

The page links a project page (https://kaiknower.github.io/mobiagent) and a GitHub repo. It does not give install commands, weights, a licence or compute figures.

Adopting the method requires:

• a π0.5 base model
• a VLM for the planner, critic and curator (the page does not name which one)
• startup demonstrations to seed the skill library
• multi-camera observations

Limits

• Truncated text. The fetched text stops partway through the method section. The ablation tables, experiment setup and any limitations section the authors wrote are missing.
• Missing parameters. Values for C, L and n_max are not given, and neither are the identity and size of the VLMs.
• Small evaluation counts. The real-robot success rates move in 2.5-point steps (32.5% → 57.5%), and RoboCasa in 2.5-point steps as well. That suggests small trial counts, though the page does not state the trial counts.
• Dependence on the curator. Whether the self-improvement gains hold up depends on how accurately the VLM curator verifies clips. The page does not report the curator's accuracy.
• Different baseline type. The comparison against π0.5-TA sets a task-specific competition model against an agentic system that uses extra VLM calls at inference time. The page does not report latency or cost for those calls.