Add PhysicalAgent

[keon]

Paper Details

• Title: PhysicalAgent: Towards General Cognitive Robotics with Foundation World Models
• Venue: arXiv, Sep 2025
• Paper: https://arxiv.org/abs/2509.13903

Summary

A general framework for cognitive robotics using a Perceive → Plan → Reason → Act pipeline. Perception and reasoning use VLMs for structured, constraint-aware plans; action generation uses a diffusion-based foundation world model to synthesize future trajectory videos; a lightweight embodiment-specific policy maps video actions to motor commands. Evaluated on mobile navigation and long-horizon manipulation, achieving up to 83% success on human-familiar tasks. Open-sourced, embodiment-agnostic design bridges foundation-model reasoning and physical execution.

Suggested Section

Embodied World Models or Reasoning & Planning

Checklist

• Add paper entry to README
• Verify links are working

[keon]

Research Verdict: ❌ SKIP

Reasons:

• Only 1 citation after 6 months (submitted Sep 2025)
• No top-venue acceptance (arXiv preprint only)
• 20-30% first-attempt success rate is notably low; relies on iterative retries to reach 80%
• List already covers this design space well:
  • Video Language Planning (ICLR 2024) for video-based planning
  • Inner Monologue (CoRL 2022) for closed-loop language feedback
  • DoReMi/CoPAL for error detection & recovery
• No code released despite stated intent
• Authors from Skoltech — recognized but not top-tier for this subfield

Recommendation: Close this issue. Reconsider if accepted at a top venue or gains meaningful citations.