Demonstration-Calibrated Port-Hamiltonian Retuning for Manipulation Policies

Yulong Yang; Fan Wu; Christine Allen-Blanchette; Amit Chakraborty

cs.ROOct 5, 2026

Demonstration-Calibrated Port-Hamiltonian Retuning for Manipulation Policies

Authors: Yulong Yang, Fan Wu, Christine Allen-Blanchette, Amit Chakraborty

Organizations: Princeton University · Arizona State University · Siemens Foundation Technologies

Abstract

Diffusion and VLA policies for manipulation are often deployed through downstream impedance controllers. The stiffness and damping gains of these controllers affect task success, yet are commonly inherited from data collection rather than selected for the deployed policy. Although empirical gain sweeps can improve performance, they require repeated evaluation rollouts. We introduce PHRetune, an offline method that derives controller gains for a frozen policy without evaluation rollouts or gain search. Our approach learns a port-Hamiltonian model from demonstrations to estimate the effort and energy associated with the policy's predicted actions. The policy is applied to recorded demonstration observations, and its predictions are assessed against demonstration-derived effort and energy budgets. From this comparison, we derive a single gain scale in closed form, adjusting the downstream controller while preserving the policy and its action representation. The gains are fixed before evaluation, without requiring a prior manipulator model, task rewards, policy retraining, or additional runtime computation. Across LIBERO suites, PHRetune improves Diffusion Policy success by up to 9.4 percentage points, with the derived gains achieving the highest observed success rates in empirical gain sweeps. On all four real-world manipulation tasks, PHRetuned Diffusion Policy outperforms the nominal policy, alternative gain-tuning methods, and a policy-retraining baseline. The same procedure improves success with SmolVLA and OpenVLA-OFT on every task, while reducing acceleration and jerk for both VLA backbones.

Figures & tables

Backbone	Model	Goal	Obj.	Spa.
Diffusion Policy	Inv- $\Sigma$	$0.229$	$0.232$	$0.244$
	DiffTune	$1.123$	$1.140$	$1.131$
	PHRetune	$0.699$	$0.461$	$0.531$

Table 1 : Simulation tuned gains. We report tuned

\beta

scaling for PHRetune, Inv-

\Sigma

, and DiffTune. Robosuite defaults

k_{p}=150

with

k_{d}

coupled via a fixed

1

damping ratio. The equivalent

\beta

values for Inv-

\Sigma

and DiffTune are calculated after their respective derivation.

Demos	Model	Goal	Obj.	Spa.	Total
$20$	Diffusion Policy	$84.6$	$76.7$	$74.3$	$78.5$
	Inv- $\Sigma$	$77.1$	$\underline{\smash{77.5}}$	$\underline{\smash{81.7}}$	$\underline{\smash{78.8}}$
	DiffTune	$80.9$	$71.5$	$67.9$	$73.4$
	MtG-SM2M	$\underline{\smash{86.1}}$	$66.7$	$80.9$	$77.9$
	PHRetune	$\mathbf{89.0}$	$\mathbf{86.4}$	$\mathbf{88.3}$	$\mathbf{87.9}$
$50$	Diffusion Policy	$\underline{\smash{92.2}}$	$\underline{\smash{86.1}}$	$88.7$	$\underline{\smash{89.0}}$

Table 2 : Diffusion Policy success rate. We report the single-task success rate on LIBERO-goal, -object, and -spatial subsets trained on 20 and 50 demos. PHRetuned Diffusion Policy achieves the highest success rate compared to baseline methods. Notably, PHRetuned Diffusion Policy trained on 20 demos approaches the performance of vanilla Diffusion Policy trained on the full 50.

Figure 1 : LIBERO gain sweep. We report the success rate over three seeds for the (a) goal, (b) object, and (c) spatial partition of LIBERO by uniformly sweeping the gain scale. The success rate for PHRetune and baselines are highlighted. MtG is omitted from the curve and

\beta

values for Inv-

\Sigma

and DiffTune are calculated post-hoc using derived gain values.

Mode		Goal	Obj.	Spa.	Total
Close Loop	$\beta^{*}$	$0.646$	$0.460$	$0.546$	$88.0$
Close Loop	SR	$89.7$	$86.3$	$88.1$	$88.0$
Shadow Mode	$\beta^{*}$	$0.699$	$0.461$	$0.531$	$87.9$
Shadow Mode	SR	$89.0$	$86.4$	$88.3$	$87.9$

Table 3 : Gain derivation in shadow mode and in closed loop. Energy-based gain scale

\beta^{*}

and success rate of the 20-demonstration policies per LIBERO suite. Shadow mode scores the frozen policy’s predicted chunks on the demonstration frames alone. Close Loop scores the chunks it executes over simulator rollouts. Both routes recover nearly the same gains and overall success rate.

Figure 2 : Real-world manipulation tasks. (a) Placing a hat on a stand, (b) transferring salt-and-pepper mills into a tray, (c) placing a camera in its fitted holder, and (d) grasping an eraser and wiping a whiteboard.

Task	Task Prompt
Hat	Place the hat on the stand
S&P	Put the salt and pepper shakers in the box
Camera	Place the camera into the slot in the box
Eraser	Pick up the eraser and erase the whiteboard

Table 4 : Real-world task prompts. We report the task prompt for all four tasks. Real-world diffusion policy receives no language conditioning; SmolVLA receives language conditioning for single-task policies; and OpenVLA-OFT receives language conditioning for all four tasks.

Backbone	Model	Hat	S&P	Camera	Eraser
Diffusion Policy	Inv- $\Sigma$	$0.217$	$0.178$	$0.163$	$0.166$
	DiffTune	$1.109$	$1.068$	$1.098$	$1.104$
	PHRetune	$0.696$	$0.461$	$0.707$	$0.859$
SmolVLA	PHRetune	$0.667$	$0.460$	$0.513$	$0.630$
OpenVLA-OFT	PHRetune	$0.731$	$0.492$	$0.476$	$0.570$

Table 5 : Real-world tuned gains. We report the derived

\beta^{*}

for Diffusion Policy, SmolVLA, and OpenVLA-OFT.

Model	Hat	S&P	Camera	Eraser	Overall
Tele-Op	1.00	0.87	0.87	0.91	0.91
Diffusion Policy	$\underline{\smash{0.6}}$	$\underline{\smash{0.4}}$	$\underline{\smash{0.5}}$	$\underline{\smash{0.5}}$	$\underline{\smash{0.50}}$
Inv- $\Sigma$	$0.2$	$0.1$	$0.0$	$0.0$	$0.08$
DiffTune	$0.5$	$0.3$	$\underline{\smash{0.5}}$	$0.2$	$0.38$
MtG-SM2M	$\underline{\smash{0.6}}$	$\underline{\smash{0.4}}$	$0.4$	$0.4$	$0.45$
PHRetune	$\boldsymbol{0.8}$	$\boldsymbol{0.6}$	$\boldsymbol{0.7}$	$\boldsymbol{0.8}$	$\boldsymbol{0.73}$

Table 6 : Real-world manipulation success rate with Diffusion Policy. We report per-task success rate and their mean across the four tasks. PHRetuned Diffusion Policy outperforms the nominal policy and baseline methods on every task. Tele-Op provides a human teleoperation reference.

Model	Hat	S&P	Camera	Eraser	Overall
SmolVLA	$0.7$	$0.5$	$0.4$	$0.6$	$0.58$
DiffTune	0.7	0.3	0.3	0.5	0.45
PHRetune	$\boldsymbol{0.9}$	$\boldsymbol{0.7}$	$\boldsymbol{0.7}$	$\boldsymbol{0.9}$	$\boldsymbol{0.80}$

Table 7 : Real-world manipulation success rate with SmolVLA. We report per-task success rate across the four tasks. PHRetuned SmolVLA improves success over the nominal policy and DiffTune on every task.

Metric	SmolVLA	PHRetune	Tele-Op
EE Jerk RMS	$22.4\pm 4.8$	$11.7\pm 4.3$ $\left(\downarrow 47.7\%\right)$	$9.3\pm 1.2$
EE Jerk Peak	$77.2\pm 17.4$	$39.3\pm 20.1$ $\left(\downarrow 49.1\%\right)$	$31.5\pm 9.0$
EE Accel. RMS	$0.93\pm 0.23$	$0.52\pm 0.17$ $\left(\downarrow 44.1\%\right)$	$0.43\pm 0.07$
EE Accel. Peak	$3.09\pm 0.58$	$1.65\pm 0.58$ $\left(\downarrow 46.6\%\right)$	$1.39\pm 0.48$
Joint Accel. RMS	$7.91\pm 3.77$	$3.34\pm 1.74$ $\left(\downarrow 57.8\%\right)$	$2.30\pm 0.34$
Time to Grasp	$3.66\pm 0.68$	$3.66\pm 0.15$	$2.99\pm 0.48$

Table 8 : Real-world motion dynamics with VLAs. We report end-effector and joint motion metrics, time to grasp, controller gain, and success rate for nominal and PHRetuned SmolVLA, PHRetuned OpenVLA-OFT over ten inference rollouts of the hat task. PHRetune reduces acceleration and jerk toward teleoperation levels with little increase to the time to grasp. Failures do not contribute to the reported metrics. End effector (EE) jerk is measured in m s -3 , EE accel. in m s -2 , joint accel. in rad s -2 , and time to grasp in seconds.

Model	Hat	S&P	Camera	Eraser	Overall
OpenVLA-OFT	$0.8$	$0.4$	$0.5$	$0.7$	$0.60$
DiffTune	0.8	0.3	0.4	0.5	0.50
PHRetune	$\boldsymbol{1.0}$	$\boldsymbol{0.6}$	$\boldsymbol{0.6}$	$\boldsymbol{0.9}$	$\boldsymbol{0.78}$

Table 9 : Real-world manipulation success rate with OpenVLA-OFT. We report per-task success rate across the four tasks. PHRetuned OpenVLA-OFT improves success over the nominal policy and DiffTune on every task.

Explore similar work

PH-Dreamer: A Physics-Driven World Model via Port-Hamiltonian Generative Dynamics
May 18, 2026Xueyu Luan, Chenwei ShiWorld Models Hamiltonian
School of Electronic and Information Engineering Tongji University Shanghai 200092, China · Shanghai Research Institute for Intelligent Autonomous Systems Tongji University Shanghai 200092, China
Co-Learning Port-Hamiltonian Systems and Optimal Energy-Shaping Control
Apr 28, 2026Ankur Kamboj, Biswadip Dey, Vaibhav SrivastavaLearning-Based Control Active Flow Control
Electrical and Computer Engineering, Michigan State University, East Lansing, MI 48824 · Meta Reality Labs, Redmond, WA 98052
Zero-shot Transfer of Reinforcement Learning Control Policies for the Swing-Up and Stabilization of a Cart-Pole System
Jun 20, 2026Nikki Xu, Hien TranReinforcement Learning Control Generative Control Policies
Department of Mathematics, Box 8205, NC State University, Raleigh, NC 27695

May 18, 2026, cs.LGJ/K move · Enter open · S save

PH-Dreamer: A Physics-Driven World Model via Port-Hamiltonian Generative Dynamics

Xueyu Luan, Chenwei Shi

School of Electronic and Information Engineering Tongji University Shanghai 200092, China · Shanghai Research Institute for Intelligent Autonomous Systems Tongji University Shanghai 200092, China

World Models Hamiltonian Dream Imagination Recurrent State Chaos

Open paper AskPDF arXiv

World models built on recurrent state space architectures enable efficient latent imagination, yet remain physically unstructured, producing dynamics that violate conservation and dissipative principles. We introduce a unified Port-Hamiltonian framework that remedies this through three synergistic mechanisms. First, we embed implicit physical priors into recurrent transitions by modeling projected latent evolution as action controlled energy routing governed by flow and dissipation, biasing the projected PH phase space toward a more compact and physically structured representation. Second, we develop a kinematics aware energy world model that estimates the Hamiltonian and power balance from proprioceptive observations, providing an explicit physical signal for thermodynamic reasoning. Third, leveraging these energy gradients, we establish an energy guided Actor-Critic that uses Lagrangian multipliers to regularize policy optimization toward lower energy and smoother control. Across visual control benchmarks, this paradigm not only attains superior asymptotic returns but also elevates internal simulator fidelity by establishing a tighter, lower variance alignment between imagined and real rewards, all while reducing latent phase space volume by 4.18-8.41%, energy consumption by up to 7.80%, and mean squared jerk by up to 9.38%.

Abstract

Explore similar work

PH-Dreamer: A Physics-Driven World Model via Port-Hamiltonian Generative Dynamics

Co-Learning Port-Hamiltonian Systems and Optimal Energy-Shaping Control

Zero-shot Transfer of Reinforcement Learning Control Policies for the Swing-Up and Stabilization of a Cart-Pole System