💻 computer science

VOFA: Visual Object Goal Pushing with Force-Adaptive Control for Humanoids

This paper presents VOFA, a hierarchical humanoid loco-manipulation system that combines a high-level visual goal-conditioned policy with a low-level force-adaptive controller to robustly push heavy objects with unknown physical properties to arbitrary goals using only onboard egocentric perception.

Original authors: Zichao Hu, Zifan Xu, Dongsik Chang, He Yin, Linh Tran, Roberto Martín-Martín, Peter Stone, Jingyu Qiao, Joydeep Biswas

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zichao Hu, Zifan Xu, Dongsik Chang, He Yin, Linh Tran, Roberto Martín-Martín, Peter Stone, Jingyu Qiao, Joydeep Biswas

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a humanoid robot named "Booster T1" standing in a warehouse. Its job is to push a heavy, mysterious box across the floor to a specific spot. The catch? The robot doesn't know how heavy the box is, it doesn't know if the floor is slippery or sticky, and it can't see the box perfectly because its camera is a bit fuzzy.

This paper introduces VOFA (Visual Object–Goal Pushing with Force-Adaptive Control), a "brain and body" system that teaches the robot how to push these boxes successfully, even when things go wrong.

Here is how VOFA works, broken down into simple concepts:

1. The Two-Brain System (The Hierarchy)

VOFA uses a two-level team approach, like a General and a Soldier:

The General (High-Level Policy): This is the "visionary" part. It looks at the camera feed (which is a bit noisy, like a bad video call) and the goal location. It decides where the robot should go and how to position itself relative to the box. It doesn't worry about the tiny muscle twitches; it just gives big commands like "Move forward" or "Turn left."
The Soldier (Low-Level Controller): This is the "muscle" part. It takes the General's big commands and figures out exactly how to move every joint in the robot's body. Crucially, this part is Force-Adaptive. Think of it like a person pushing a shopping cart that suddenly gets filled with bricks. If the cart gets heavy, the person instinctively leans harder and adjusts their footing so they don't fall over. The Soldier does the same thing automatically, adjusting to the box's weight and the floor's friction without needing to be told.

2. The Training Camp (Teacher-Student Method)

Teaching a robot to do this is hard because you can't easily give it a "perfect view" of the world during training. So, the researchers used a Teacher-Student method:

The Teacher: First, they trained a "Teacher" robot that had superpowers. It could see the exact weight of the box, the exact friction of the floor, and the perfect position of the robot's joints. It learned how to push the box perfectly using this "privileged" information.
The Student: Then, they trained a "Student" robot that only had a regular camera and basic sensors (no superpowers). The Student watched the Teacher and tried to copy its moves. To make the Student tougher, the researchers added "visual noise" during training—like blurring the camera or adding static—so the Student learned to push the box even when the view was messy, just like it would be in the real world.

3. The "Align First" Trick

One of the biggest challenges is that if the robot tries to push the box when it's standing on the wrong side, it will fail. The researchers added a special "reward" (like a gold star) for the robot to reposition itself first.

Imagine trying to push a heavy couch. If you stand on the side, you can't move it forward. You have to walk around to the back first. VOFA learned this trick on its own. It waits until it is standing in the perfect spot behind the box relative to the goal before it starts pushing. This prevents the robot from making a "premature contact" and getting stuck.

4. Real-World Results

The team tested this on the real Booster T1 robot. Here is what happened:

Heavy Lifting: The robot successfully pushed boxes weighing up to 17 kg (about 37 lbs). That is more than half the weight of the robot itself! The robot wasn't even trained on boxes this heavy in the simulation, yet it figured it out on the fly.
Different Angles: Whether the goal was in front, to the side, or even behind the robot, VOFA succeeded over 80% of the time in the real world.
Recovery: If someone kicked the box sideways while the robot was pushing it, the robot didn't panic. It stopped, looked at where the box went, walked around to get back in line, and continued pushing. It operates in a closed loop, meaning it constantly checks its surroundings and corrects its path, rather than just following a pre-planned script.

Summary

VOFA is a system that lets a humanoid robot push heavy, unknown objects to specific targets using only a camera. It combines a smart "visionary" planner with a "muscle" that instinctively adjusts to weight and friction. By training a student robot to mimic a super-powered teacher while dealing with messy camera data, the system learned to push heavy boxes, recover from bumps, and navigate complex angles without falling over.

Technical Summary: VOFA – Visual Object Goal Pushing with Force-Adaptive Control for Humanoids

Problem Statement

The paper addresses the challenge of enabling humanoid robots to perform goal-directed object pushing (loco-manipulation) using only onboard egocentric perception. This capability is critical for real-world applications such as warehouse inventory reorganization. The system must operate robustly under three primary constraints:

Unknown Physical Properties: The robot must handle objects with varying and unknown masses and ground friction coefficients.
Noisy Onboard Perception: The system cannot rely on privileged object-state information (e.g., ground-truth object position or velocity) and must function using noisy depth images and proprioceptive data.
Closed-Loop Adaptability: The robot must adapt in real-time to diverse object-goal configurations, including cases where the robot must reposition itself relative to the object and goal before initiating contact.

Existing solutions often fail to meet these criteria simultaneously, either relying on privileged sensing, lacking robustness to physical variations, or failing to adapt to different goal configurations without manual stage definitions.

Methodology

The authors propose VOFA (Visual Object–Goal Pushing with Force-Adaptive Control), a hierarchical, two-level system designed for the Booster T1 humanoid robot.

1. Hierarchical Architecture

High-Level Visuomotor Policy: This policy processes onboard observations (proprioception and depth images) and a desired goal position to generate high-level commands. It operates at 50Hz.
Low-Level Force-Adaptive Controller: This controller translates high-level commands into joint-level actions. It is based on the FALCON framework, explicitly designed to handle substantial and unknown external forces at the end-effectors, ensuring stability during interaction with objects of varying mass.

2. Training Strategy: Teacher-Student Distillation

To overcome the challenge of learning from noisy visual inputs, VOFA employs a Teacher-Student framework:

Teacher Policy: Trained using Proximal Policy Optimization (PPO) with privileged information (e.g., exact object orientation, base linear velocity, and object state). This policy learns the optimal behavior for goal-conditioned pushing.
Student Policy: Distilled from the teacher using DAgger (Dataset Aggregation). The student replaces privileged inputs with onboard proprioceptive history and depth images.
Visual Augmentation: To bridge the sim-to-real gap, the student training incorporates visual randomizations, including far-plane depth perturbation, correlated depth noise, and pixel dropout, to simulate sensor artifacts and background clutter.

3. Reward Design

The training utilizes a composite reward function with four key terms:

Reaching Reward ( $R_{reach}$ ): Encourages both end-effectors to approach the object.
Pushing Reward ( $R_{push}$ ): Rewards the reduction of the distance between the object and the goal.
Head Tracking Reward ( $R_{track}$ ): Encourages the robot to visually track the object with its camera.
Object-Goal Alignment Reward ( $R_{align}$ ): A novel component that encourages the robot to reposition itself to the side of the object opposite the goal before pushing. This eliminates the need for manually defined training stages and promotes long-horizon planning over myopic contact strategies.

4. Simulation and Domain Randomization

Training is conducted in Isaac Gym with extensive randomization of object mass (1–8 kg), friction, and initial object-goal placements. To address the sim-to-real gap in vision, the system randomizes camera extrinsics (translations and rotations) and emulates sensor noise patterns observed in real-world depth cameras (ZED 2i).

Key Contributions

Visual Goal-Conditioned System: Introduction of VOFA, a system capable of closed-loop, goal-directed pushing across diverse object-goal configurations using only onboard visual perception.
Robustness to Unknown Masses: Demonstration of a system that successfully pushes objects with diverse and previously unseen masses, including loads up to 17 kg (exceeding half the robot's body weight), without retraining on those specific masses.
Validation on Real Hardware: Successful deployment on the Booster T1 humanoid robot, achieving high success rates in both simulation and real-world experiments across varying scene configurations and physical properties.

Experimental Results

The system was evaluated in both large-scale simulation and real-world trials on the Booster T1.

Success Rates: VOFA achieved over 90% success in simulation and over 80% success in real-world trials.
Mass Robustness: In real-world experiments, the system successfully pushed a 17 kg object (a 5 kg box with 12 kg of added weight), despite being trained on masses up to 8 kg.
Configuration Adaptability: The system maintained high success rates (>90% for the student policy) across front, lateral, and rear goal configurations.
Ablation Studies:
- Force-Adaptive Control: Removing the force-adaptive low-level controller resulted in a significant drop in success rates (e.g., from 91.38% to 59.32% for the student policy) and unstable behaviors like "kicking" the object.
- Alignment Reward: Removing the object-goal alignment reward led to premature contact failures, particularly in rear-goal configurations where repositioning is required.
- Visual Randomization: Training without visual randomizations led to substantial performance degradation in real-world lateral and rear scenarios due to sensitivity to background noise and domain mismatch.
Closed-Loop Capabilities: Real-world demonstrations showed the robot could recover from external lateral perturbations, compensate for off-centered center-of-mass distributions, and sequentially push objects to multiple goals without reset.

Significance and Claims

The paper claims that VOFA represents a significant step toward robust, autonomous humanoid loco-manipulation in unstructured environments. By decoupling high-level visuomotor planning from low-level force-adaptive control, the system simplifies policy training while maintaining robustness to physical variations.

The authors emphasize that the object-goal alignment reward is critical for enabling the robot to learn complex, long-horizon behaviors (repositioning before pushing) without manual staging. Furthermore, the successful transfer to real-world hardware with heavy loads (17 kg) demonstrates that the combination of force-adaptive control and visual randomization effectively mitigates the sim-to-real gap.

The work concludes by noting current limitations, specifically the focus on single-object scenarios and the reliance on an external module for relative goal positioning. Future work aims to extend the framework to multi-object environments and develop fully end-to-end reasoning for object and goal localization from raw sensory input.