OIL: An Object Interaction Layer for Humanoid Whole-Body Control

Pranay Dugar, Mohitvishnu S. Gadde, Rajesh Mangannavar, Aayam Kumar Shrestha, Alan Fern · N/A · 2026

Summary

Recent progress in learned whole-body controllers for humanoid robots has enabled robust free-space motion by tracking kinematic targets such as joint or end-effector trajectories.

Abstract Summary

Recent progress in learned whole-body controllers for humanoid robots has enabled robust free-space motion by tracking kinematic targets such as joint or end-effector trajectories. This interface provides a useful abstraction for upstream systems, including motion capture, teleoperation, or generative models, which can specify desired motion without reasoning about forces or balance. However, this abstraction has largely been limited to robot motion, requiring task-specific controllers or retraining for object interaction. We introduce OIL, an Object Interaction Layer that extends the kinematic command interface of whole-body control to include desired object motion and contact. OIL augments a standard controller with object-centric commands and translates them into coordinated whole-body behavior while preserving locomotion and balance. This design enables a single learned policy to execute non-prehensile interactions, including pushing, pulling, two-arm pick-and-place, and locomotion-coupled transport, directly from object-level kinematic targets without task-specific retraining. In simulation, we demonstrate that object-level commands improve performance over robot-centric controllers while matching task-specific policies across variations in object properties such as mass and size. We further demonstrate zero-shot generalization to novel behavior combinations and sim-to-real transfer on the Digit humanoid robot.

Key Points

Recent progress in learned whole-body controllers for humanoid robots has enabled robust free-space motion by tracking kinematic targets such as joint or end-effector trajectories.
This interface provides a useful abstraction for upstream systems, including motion capture, teleoperation, or generative models, which can specify desired motion without reasoning about forces or balance.
However, this abstraction has largely been limited to robot motion, requiring task-specific controllers or retraining for object interaction.
We introduce OIL, an Object Interaction Layer that extends the kinematic command interface of whole-body control to include desired object motion and contact.
OIL augments a standard controller with object-centric commands and translates them into coordinated whole-body behavior while preserving locomotion and balance.

Problem Formulation

We consider the problem of learning a humanoid controller capable of executing a broad distribution of object-interaction behaviors specified through kinematic directives. We focus on non-prehensile bimanual end-effector interaction, where the robot manipulates objects using coordinated whole-body motion and contact without dexterous grasping.

These behaviors include pushing, pulling, lifting, carrying, and locomotion-coupled transport, with variations in object motion, contact timing, and base movement.

Problem Formulation — OIL: An Object Interaction Layer for Humanoid Whole-Body Control — Figure 1: Overview of OIL. An object-interaction layer that augments a whole-body controller with object-level motion and contact directives, enabling a single controller to execute diverse loco-manipulation behaviors such as walking up to a box, picking it up from different heights and positions in front of the robot, walking while holding it up, placing it down at different locations, and pushing it on the surface with one or two arms. It further allows composition of these skills without any per-task finetuning.

Object-Interactive Whole-Body Control

We now describe how the directive formulation introduced in the previous section is realized in practice. Our approach builds an object-interactive controller by layering an Object-Interaction Layer (OIL) on top of an existing, well-trained whole-body controller (WBC) trained for free-space (non-interactive) motion.

In this work, we use the Masked Humanoid Controller (MHC) as the underlying WBC due to its robustness, support for partially specified kinematic directives, and demonstrated sim-to-real transfer capabilities.

Object-Interactive Whole-Body Control — OIL: An Object Interaction Layer for Humanoid Whole-Body Control — Figure 2: Architecture of the Object-Interactive Controller . A behavior generator produces an object-interaction directive, split into an interaction directive $d^{i}_{t}$ (desired object position $\hat{o}_{t}$, end-effector targets $\hat{p}^{R}_{t}$, $\hat{p}^{L}_{t}$, and contact conditions $\hat{c}^{R}_{t}$, $\hat{c}^{L}_{t}$) and a base directive $d^{b}_{t}$ (desired velocity $\hat{v}_{t}$, turn rate $\hat{\omega}_{t}$, and base orientation $\hat{b}_{t}$). OIL (blue), running at 50 Hz, maps $d^{i}_{t}$, object state (updated at 10 Hz), and robot proprioception to an arm joint directive $u_{t}$. The Masked Humanoid Controller (MHC, green), also running at 50 Hz, combines $u_{t}$ and $d^{b}_{t}$ with robot proprioception to produce PD targets $a_{t}$ for all actuated joints, which are executed by a PD controller at 2 kHz. During training, the MHC weights remain frozen and only the OIL weights are optimized.

OIL Training

We train the OIL policy $\pi^{\mathrm{OIL}}$ via reinforcement learning in the MuJoCo simulator with the Digit V3 humanoid model using Proximal Policy Optimization (PPO) ( Schulman et al. 2017 ) . The underlying whole-body controller $\pi^{\mathrm{WBC}}$ (MHC in our experiments) remains frozen throughout, so learning is confined to $\pi^{\mathrm{OIL}}$.

Below, we describe the training procedure, including episode sampling, reward design, and domain randomization. Training is episodic.

OIL Training — OIL: An Object Interaction Layer for Humanoid Whole-Body Control — Figure 3: Out-of-distribution robustness to object properties for OIL and task-specific policies. Left: success rate on bi-manual slide as object width increases beyond the training range. Right: success rate on uni-manual slide as object mass increases beyond the training range. Dashed lines indicate training cutoffs.

Results

We evaluate OIL along three axes that correspond to its design claims: (i) whether a single policy with object-level commands improves over robot-only control and matches task-specific policies trained per behavior, (ii) whether the OIL controller generalizes to object properties outside its training distribution, and (iii) whether a single controller supports zero-shot composition of novel behavior sequences.

FUTURE WORK

Future work will expand the interaction directives to include full SE(3) object orientation, different grasp modes, and force/contact targets when this information is available upstream. Another direction is to generate directives in a closed loop, allowing the system to replan when contact fails or degrades instead of being limited to a single attempt.

We also plan to incorporate visual observations as input for object sensing, reducing reliance on externally provided object state estimates.

Sources and Demonstrations

Figures are reproduced from Dugar et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.