About Me

Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy

Evaluating a frontier agent on EBench under direct robot control and with a harness that can invoke post-trained robot policies.

PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models

Cross-appearance self-alignment encourages world action models to produce consistent action representations under task-irrelevant visual changes.

InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data

A world action model combining predictive dynamics, semantic guidance, and geometric priors, pretrained on more than 20,000 hours of open data.

InternW0: A Foundational Physical World Model for Efficient Real-World Interactions

An asymmetric video-action architecture couples future visual dynamics with continuous robot control, supporting asynchronous prediction and contact-aware interaction.

InfCoiL: Coordinated Planner-Controller Learning for Closed-Loop Physics-Based Human-Object Interaction

Accepted

Make generated motion executable. InfCoiL lets a flow-matching planner and a physics-based controller improve each other through shared rollout feedback, alternating their RL updates when progress plateaus.

Exploratory, Communicative, and Deployable: Vision-Driven Embodied Agents for Open-World Mobile Manipulation

Vision-driven mobile manipulation through active exploration, visual grounding, and interactive intent disambiguation, with a shared interface for simulation and real robots.

SSRN 2026

Length-Adaptive PPO-based RLHF for Non-Uniform UAV Control

Length-adaptive proximal policy optimization with reinforcement learning from human feedback for non-uniform UAV control.

AnyBipe: An Automated End-to-End Framework for Training and Deploying Bipedal Robots Powered by Large Language Models

Language-guided reward design, reinforcement learning, and sim-to-real evaluation in one end-to-end pipeline for bipedal robots.

GBC: Generalized Behavior-Cloning Framework for Whole-Body Humanoid Imitation

An open-source pathway from human motion to humanoid action: differentiable retargeting, DAgger-MMPPO, and an Isaac Lab training workflow.

Learning Embodied Agents from Experience

In progress

How can an open model learn to act from a frontier agent’s experience? I am studying the transfer of Astra’s robot interaction experience to Qwen3.8-27B through staged SFT and reinforcement learning, connecting visual grounding, short manipulation skills, and long-horizon tasks.

ARIEL: An Agent-Ready Humanoid Behavior Model with In-Context Execution Memory

In development

ARIEL gives an agent a simple interface to whole-body control. Choose a demonstration, load it into memory, then recall the motion or steer its style with velocity and heading commands. The controller keeps executing between agent decisions.

Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents

arXiv release pending

An embodied agent learns from execution experience by refining text memory across training episodes, then uses frozen memory at evaluation without updating model weights.

ABI: Actor-Free Zero-Shot Reinforcement Learning with Behavior-Inductive Representations

Manuscript in preparation

Given an existing policy, ABI trains only a reusable guide. The guide improves task execution at deployment while the policy stays frozen, with no parameter updates or task-specific policy post-training.