RoboReplan
RoboReplan is a tabletop planning environment for OpenEnv that targets Problem Statement 3.1 (World Modeling: Professional Tasks).
Core trick
Injects blockers, grasp slips, and mid-task instruction changes into a tabletop pick-and-place env so the LLM must replan instead of looping.
Description
RoboReplan is a tabletop planning environment for OpenEnv that targets Problem Statement 3.1 (World Modeling: Professional Tasks). LLMs often fail at long-horizon robot tasks because they can't replan — they loop or freeze when something goes wrong (blockers, grasp slips, mid-task instruction changes). RoboReplan benchmarks this failure mode and trains agents to recover: clear blockers, pick and place in the right bins, and adapt when the instruction changes. The env is deliberately hard for a small model so we can show improvement over time (0% → 78% success with SFT + GRPO). Four domain skins (Default, Pharmacy, Warehouse, Lab) ground the same mechanics in professional scenarios. Built with OpenEnv 0.2.1, deployed on HF Spaces, trained with HF TRL (GRPO) in Colab
README excerpt
> **DEPRECATED**: Superseded by [physical-grounding-lab](https://github.com/jwalin-shah/physical-grounding-lab) as of 2026-05-19. This repo is archived for historical reference. # RoboReplan — Tabletop Robot Planning Environment **Hackathon Problem Statement 3.1 — World Modeling: Professional Tasks** > Agents must maintain consistent internal state, update beliefs based on outcomes, > and orchestrate multi-step workflows in a dynamic, partially observable world. --- ## The Problem LLMs fail at long-horizon robotic tasks not because they can't move, but because **they can't replan**. When a grasp slips, when a blocker appears, when the instruction changes mid-task — the model freezes, repeats the same failing action, or abandons the plan entirely. RoboReplan benchmarks exactly this failure mode and trains agents to recover from it. --- ## What RoboReplan Tests A tabletop scene with 2–5 objects and 1–2 target bins. The agent receives a natural-language instruction and must: - **Decompose** the instruction into an ordered plan - **Handle blockers** — clear whatever is in the way before picking the target - **Replan after failures** — grasp slips, partial clears, and perception noise require retry logic - **Respect constraints** — fragile first, heavy last, urgent first - **Track state** — know what's placed, what's held, what's failed, across many steps - **Adapt mid-task** — instructions can change at step 6 or 12; the agent must update its plan ### Professional Tas