Physics-Informed Causal Reasoning for Nonstationary Robot Dynamics
Authors: Jungjin Lee, Se Jin Kim, Sung Kwon On, Insoon Yang
Organizations: Department of Electrical and Computer Engineering, Seoul National University, Seoul 08826, Korea · Graduate School of AI, Seoul National University, Seoul 08826, Korea
Reasoning about what an action will cause requires internalizing how a system behaves. As conditions seldom stay fixed, this must often happen within a few interactions. Yet systems that look identical in a snapshot can respond differently to the same action, depending on physical conditions the snapshot hides. Context-based models capture this separation by summarizing recent transitions into a latent context, but do not specify how that context should change the equations of motion. Physics-informed neural networks build those equations into the model, but describe only a single system and cannot recognize when it has changed. We present Physics-Informed Causal Reasoning (PICR), which bridges the two. Only a context, re-inferred from recent interactions, is updated; it conditions a fixed physics-informed network that determines the robot's inertia and potential energy, so the robot adapts to new physical conditions from its own experience without retraining or a hand-specified robot model. On a Franka Panda arm in NVIDIA Isaac Sim, PICR predicts motion under changing and unseen payloads, reducing error on the latter by 1.6-3.8x over baselines, and extends to counterfactual queries under alternative torques. Our analysis shows that the context, inferred solely from interaction, encodes payload physical properties, tracks their changes, and governs the arm's response. The gravity and inertia terms induced by the context further demonstrate its physical validity, each agreeing with the simulator at the level of individual Lagrangian components.
Figures & tables
Figure 1: PICR overview. A context inferred from recent interactions conditions a fixed Lagrangian network via FiLM, mirroring how unobserved payload properties shape the robot’s rigid-body dynamics.
Figure 2: Prediction under a payload ramp (0.1 → 4 kg). (a) Figure-eight task in Isaac Sim. (b) Top: predicted (coloured) vs. true (grey) cube. Bottom: 300 ms center error over all windows; grey square: cube (4.8 cm).
Figure 3: Counterfactual torque interventions. (a) Intervention configuration. (b) Simulated vs. PICR motion 300 ms after a ± 20 N m offset to joint 2. (c) Projected gain ρ vs. ∣Δτ∣ (joint 2); shading: PICR IQR.
Metric
Setting
PICR
RWM
CaDM
CaDeLaC
DeLaN
Neural-Fly
ADD (cm) ↓
Seen (0.1–2 kg)
4.93
7.45
6.06
8.75
6.96
8.05
Unseen (2–4 kg)
6.76
11.15
11.40
21.41
25.78
22.27
Gain ρ→1
Joints 1–3 (mean)
0.90
0.75
0.65
0.17
0.35
0.84
RMSE (N m) ↓
Gravity g2
2.15
n/a
n/a
11.29
9.99
n/a
Inertial [Hq¨]2
2.31
n/a
n/a
12.49
5.83
n/a
Coriolis c2
1.59
n/a
n/a
2.15
1.67
n/a
Table 1: Prediction, counterfactual and component evaluation (mean over three seeds). ADD: figure-eight payload sweep, seen (0.1–2 kg) vs. unseen (2–4 kg) range. Gain: median acceleration gain under ± 10–40 N m torque offsets. RMSE: joint-2 torque terms vs. simulator on a 3 kg carry. n/a: no explicit H , c , g . Best in bold.
Figure 4: Physical information in the inferred context. (a) PCA of z coloured by payload mass. (b) Context substitution on a 3 kg window. (c) PC1 after payload lift. (d) PC1 as the payload ramps between 0.1 and 4 kg.
Figure 5: Learned Lagrangian terms follow the payload. Joint-2 terms vs. simulator (dashed). (a) 3 kg carry (unseen). (b) Payload jump 0.1 → 2 kg (median of 472 episodes; band: PICR IQR). (c) Across payload masses (shaded: unseen).
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Context from history
Output
Parameters
Optimizer
PICR
MLP encoder, z∈R4 , FiLM
q¨ˉ (Lagrangian)
6,545,957
Adam 3×10−4 , cosine
RWM
GRU state, 256-d
next state
956,184
Adam 10−3 , clip 1
CaDM
MLP encoder, e∈R10
state change
231,008 †
Adam 10−3
CaDeLaC
LSTM encoder, z∈R10 , concat
q¨ˉ (Lagrangian)
6,719
Adam 10−3 , clip 1
DeLaN
none
q¨ˉ (Lagrangian)
12,125
Adam 10−4
Neural-Fly
least-squares weights a∈R3×7
q¨ˉ
133,719 †
Adam 5×10−4
Appendix
Table 2: Baselines. Parameters count the prediction path. † CaDM’s backward core (129,814) and Neural-Fly’s discriminator (248) are used only in training.
Figure-eight ADD [cm] ↓
0.1–2 kg
2–4 kg
g2 / [Hq¨]2 / c2 RMSE [N m] ↓
PICR
4.58
6.92
1.84 / 1.63 / 1.46
PICR w/o contrastive
6.07
9.78
2.11 / 2.87 / 1.15
Appendix
Table 3: Removing the contrastive term from PICR ( λcon=0 ; seed 0, final epoch). Term RMSE is for joint 2 on a 3 kg carry.
Figure-eight ADD [cm]
0.1–2 kg
2–4 kg
Gain at 20 N m
RWM
10.58
13.57
0.73
RWM + contrastive
12.48
13.21
0.55
CaDM
9.76
10.56
0.63
CaDM + contrastive
11.06
11.66
0.67
CaDeLaC
11.33
22.16
0.16
Appendix
Table 4: Adding PICR’s contrastive term to the baselines (seed 0, N=15 ). The gain is the median one-step projected acceleration gain at ∣Δτ∣=20 N m, averaged over joints 1–3.
Figure-eight ADD [cm] ↓
0.1–2 kg
2–4 kg
g2 / [Hq¨]2 / c2 RMSE [N m] ↓
FiLM (PICR)
4.93±0.54
6.76±1.28
2.15 / 2.31 / 1.59
Input concat
4.45±0.37
8.49±0.69
4.02 / 2.27 / 2.09
Shift only
4.59±0.54
8.43±0.80
3.33 / 2.29 / 1.96
Low-rank hypernet
4.26±0.35
11.75±2.63
5.85 / 1.91 / 3.02
Appendix
Table 5: Conditioning mechanisms for the Lagrangian networks. ADD is the mean ± standard deviation over three seeds; term RMSE is the seed mean. Final epoch. The FiLM row corresponds to PICR in Table 1 ; its seed 0 is the model of Tables 3 and 6 .
Figure-eight ADD [cm] ↓
Parameters
0.1–2 kg
2–4 kg
PICR
6.55M
4.58
6.92
PICR, width 256
0.56M
4.77±0.19
8.66±1.52
RWM
0.96M
7.09
11.63
CaDM
0.23M
6.79
10.90
CaDeLaC
6.7k
9.07
24.75
Appendix
Table 6: Lagrangian models at a common size of about 0.5M parameters. Only layer widths differ from the rows of Table 1 (seed 0); the narrow PICR is the mean ± standard deviation over three seeds. Final epoch.
Set
Motion and control
Payload condition
Size and use
Training
Joint excitation; PID with OU torque perturbation
0.1 – 2.0kg , 8 levels; 53 CoM lattice
318 training / 40 validation episodes
Payload sweep
PPO figure-eight tracking
Mass ramp from 0.1 to 4.0kg
33 episodes / 32,967 windows
Torque intervention
PPO replay with torque offsets
0.1 – 2.0kg ; grasp offset along hand x
1,887 of 5,616 pairs
Appendix
Table 7: Datasets used in this work. Only the training split is used to fit the models. Each window contains a 12-step history and a 15-step prediction horizon.
Figure 6: (a) Payload center-of-mass (CoM) variation in the training data. The CoM of the 4.8cm cube is offset independently along each axis to one of {0,±9.8,±19.6}mm , giving a 53=125 -point lattice (orange: cube center). (b–d) The latent z of every training window, projected onto the same PC1–PC2 plane as Fig. 4 (a) and coloured by the CoM offset along x (b), y (c) and z (d). Grey points are windows from the held-out 2kg4kg payloads, whose CoM is fixed at the cube center.
Reinforcement learning (RL) has achieved strong performance in robotic control; however, state-of-the-art policy learning methods, such as actor-critic methods, still suffer from high sample complexity and often produce physically inconsistent actions. This limitation stems from neural policies implicitly rediscovering complex physics from data alone, despite accurate dynamics models being readily available in simulators. In this paper, we introduce a novel physics-informed RL framework, called PIPER, that seamlessly integrates physical constraints directly into neural policy optimization with analytical soft physics constraints. At the core of our method is the integration of a differentiable Lagrangian residual as a regularization term within the actor's objective. This residual, extracted from a robot's simulator description, subtly biases policy updates towards dynamically consistent solutions. Crucially, this physics integration is realized through an additional loss term during policy optimization, requiring no alterations to existing simulators or core RL algorithms. Extensive experiments demonstrate that our method significantly improves learning efficiency, stability, and control accuracy, establishing a new paradigm for efficient and physically consistent robotic control.
We present PhysCaP, a Physics-Informed Code-as-Policy agent system for active perception in robotic manipulation. While vision-language-action policies excel at imitating demonstrations, they rely on passive observation and fail to infer latent physical properties critical for manipulation. PhysCaP augments code-as-policy frameworks with a physics-informed exploration layer that enables explicit information-seeking through interaction. Our method introduces training-free physical property extraction modules that estimate object mass and stiffness from robot proprioception without additional sensors. To balance exploration costs and the efficiency of information obtained, PhysCaP employs a multi-agent design: a Planner that decides when to explore and when to stop, and a Prioritizer that filters implausible interactions and ranks the remainder using a heuristic priority score, enabling efficient, targeted exploration. We evaluate PhysCaP on three real-world tabletop manipulation tasks and a simulated task in LIBERO. The results show that existing passive and naive interactive baselines either fail when physical properties are hidden or over-explore, whereas PhysCaP achieves comparable performance with fewer interactions and reduced execution time. Ablation studies further validate the effectiveness of the proposed physical property extraction modules. Project page: https://physcap.github.io
Chen-Yu Lin, Jing-Wen Chen, Hsueh-En Chang +8
National Taiwan University · NVIDIA Research · National Yang Ming Chiao Tung University
Static shape estimation of co-manipulative continuum robots (CCRs) is challenging because the continuum arms and manipulated flexible object form a closed chain that must satisfy both static equilibrium and geometric loop-closure constraints. This paper presents a constraint-aware physics-informed neural network (PINN) for static shape estimation of a tendon-driven CCR modeled using the geometric variable strain formulation. The proposed method incorporates a projected static equilibrium residual and a configuration-level geometric residual to enforce the governing mechanics and closed-chain geometry. In simulation, the PINN is compared with a purely data-driven artificial neural network (ANN) under limited and noisy training data. With 140 samples and 50% label noise, the PINN reduces the relative configuration error, equilibrium residual, and closed-chain residual by 67.88%, 67.35%, and 88.06%, respectively. Using the full dataset, the PINN achieves 0.1597% relative configuration error with an inference time of 0.1773 ms, compared with 17.97 s for an iterative nonlinear solver. Experimental fine-tuning reduces the marker RMSE from 2.657 mm to 0.497 mm and increases R2 from -0.788 to 0.937. These results demonstrate accurate, physically consistent, and computationally efficient static shape estimation of closed-chain CCRs.
Rana Danesh, Pari Qarehdaghi, Farrokh Janabi-Sharifi