Research map / Reinforcement learning / Visual and representation RL

LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Dexterous manipulation transfer

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces LIBERO-PeRM, a simulated benchmark of five procedurally generated task suites (312 tasks, 40.8K rollouts) for evaluating personalized robotic manipulation via language-expressible motion, skill, and user-profile preferences, and benchmarks VLA policies under zero-shot and LoRA fine-tuning.

Problem

Personalized robotic manipulation lacks a unified formulation and controlled benchmark that separates task completion from user preference satisfaction, especially for fine-grained manipulation preferences and multi-preference conflicts.

Approach

Treat user preferences as a variable independent of the task goal, define a taxonomy (motion, skill, user profile) of observable preferences in LIBERO, and procedurally generate paired preference-conditioned and no-preference demonstrations with automatic preference-satisfaction metrics.

Key results

• Zero-shot, pi0.5 reaches 97.6% no-preference SR but only 24.38% (motion) and 25.00% (skill) preference SA, showing it ignores preferences.
• Mix-LoRA retains 91.00% no-pref SR vs 52.33% for Seq-LoRA, but drops Motion-preference SR from 90.00% to 66.88% relative to preference-only LoRA.
• Preference-preference conflict drops SR/SA to 66.00%/47.75%; composition reaches 25.0% SR/SA and decomposition yields 0.0% SA despite 37.5% SR.

Surprising findings

• Task success does not imply preference satisfaction: decomposition gets 37.5% SR but 0.0% SA.
• Mixed training improves no-preference retention but weakens motion-preference learning relative to preference-only LoRA.

Limitations

• Preferences are predefined, language-expressible, and automatically measurable in simulation; excludes implicit inference, interaction history and open-ended preferences.
• Generated demonstrations may carry pipeline and skill-library biases; human validation is a small 15-person survey with no real-robot or longitudinal study.
• Composition covers only two preferences and the conflict suite does not test conflict detection or clarification.

Experimental setup

Datasets: LIBERO

Benchmarks: LIBERO-SinglePref LIBERO-PrefScale LIBERO-PrefConflict LIBERO-PrefCompose LIBERO-ComprePref

Metrics: task success rate (SR) preference satisfaction (SA)

Baselines: OpenVLA-OFT pi0-fast GR00T-N1.6 SmolVLA UniVLA

Builds on: LIBERO OMPL pi0.5

Closest papers by meaning

• Promptable Behaviors: Personalizing Multi-Objective Rewards from Human Preferences (CVPR 2024 · similarity 0.91)
• SYNAPSE: SYmbolic Neural-Aided Preference Synthesis Engine (AAAI 2025 · similarity 0.88)
• Human-assisted Robotic Policy Refinement via Action Preference Optimization (NeurIPS 2025 · similarity 0.88)
• PEARL: Zero-shot Cross-task Preference Alignment and Robust Reward Learning for Robotic Manipulation (ICML 2024 · similarity 0.87)
• Predictive Preference Learning from Human Interventions (NeurIPS 2025 · similarity 0.87)
• Direct Post-Training Preference Alignment for Multi-Agent Motion Generation Model Using Implicit Feedback from Pre-training Demonstrations (ICLR 2025 · similarity 0.87)