libero

🎁 A collection of utilities for LeRobot.

Robot-agent harness with a CLI and local Web console for LIBERO short evaluation

A Python package to use FPGA development tools programmatically.

A counterfactual battery for instruction following in vision-language-action policies: change one word of the instruction, hold the scene, and record which object the arm reaches. Ships the deconfounded demonstration generators and their confounded control.

Agent + reusable skill library for LIBERO manipulation tasks (no RL / VLA / world model; no LLM calls at runtime)

Turn world-model imagination into corrective training data for robot policies — a single-GPU, falsifiable pipeline (SmolVLA · LIBERO · CEM · paired simulator replays).

[CoRL 25] Code for FLOWER VLA for finetuning FLOWER on CALVIN and all LIBERO environments

Fine-grained robot control with Jev, physics previews, and configurable LIBERO tasks.

Official code for AffordanceVLA: Empowering Action Generation through Affordance-Aware Understanding.

[NeurIPS 2025] Code for BEAST Experiments on CALVIN and LIBERO.

Anchor-Align (arXiv:2607.13429): VLA finetuning that prevents behavior cloning from erasing pretrained VLM representations (catastrophic forgetting) and aligns language with actions. OOD generalization on a physical xArm7, LIBERO-PRO, LIBERO-Plus and CALVIN.

Official implementation of PhaseLoRA: control-regime-conditioned low-rank adaptation for continuous-action vision-language-action policies, built on OpenPI.

SUREFlow is a lightweight, Mamba-based vision-language-action model with 179M parameters for robot manipulation. Paper: https://arxiv.org/pdf/2607.10504

SmolWAM: a 0.88B latent world-action model for robot manipulation on edge devices. NeurIPS 2026 Workshop on Robotics World Modeling.

Backdoor training and evaluation scripts for vision-language-action models on LIBERO.

A simple, efficient, and extensible evaluation framework for robot policies across diverse benchmarks, supporting both local and remote execution

Check whether a VLA benchmark number is measuring what you think it measures — checkpoint integrity, failure taxonomy, headroom & power analysis (zero-dependency, framework-agnostic).

Automated agentic research loop for vision-language-action (VLA) robot policy self-improvement

Run and compare robot VLAs and world-action models in LIBERO and RoboCasa with reproducible local evidence.

Official code repository for SkillNet