Idea description

Replace additive neural state updates for rotations or rigid poses with a learned forced dynamical system whose configuration is updated by Lie-group multiplication. The network predicts body-frame force or acceleration in the Lie algebra, while the exponential map guarantees that every predicted configuration remains on SO(3) or SE(3).

Formulas

$$\frac{d}{dt}\left(\frac{\partial L}{\partial\dot{q}}\right)-\frac{\partial L}{\partial q}=F.$$

$$\dot g=T_eL_g(\xi),\qquad g\in G,\ \xi\in\mathfrak g=T_eG.$$

$$g_{k+1}=g_k\operatorname{Exp}(h\xi_k),\qquad \xi_{k+1}=\xi_k+h\,a_\theta(g_k,\xi_k,c_k),\qquad a_\theta(g,\xi,c)\in\mathfrak g.$$

$$\mathcal L_{\mathrm{rollout}}(\theta)=\sum_{k=1}^{H}d_G\!\left(\hat g_k,g_k\right)^2+\lambda\sum_{k=0}^{H-1}\left\|a_\theta(g_k,\xi_k,c_k)\right\|_2^2,$$

Mathematical statement

The paper starts from the forced Euler-Lagrange equation $\frac{d}{dt}(\frac{\partial L}{\partial\dot q})-\frac{\partial L}{\partial q}=F$, where $Q$ is a configuration manifold, $L:TQ\to\mathbb{R}$ is a Lagrangian, $q\in Q$ is configuration, $\dot q\in T_qQ$ is velocity, and $F:TQ\to T^*Q$ is an external covector-valued force. For a Lie-group configuration $Q=G$, represent velocity by a Lie-algebra element $\xi\in\mathfrak g=T_eG$ using $\dot g=T_eL_g(\xi)$, where $g\in G$, $e$ is the identity, and $L_g(h)=gh$ is left translation. The neural adaptation uses a learned force or acceleration map $a_\theta(g,\xi,c)\in\mathfrak g$, with context $c$ such as observations or controls, and integrates it using $g_{k+1}=g_k\operatorname{Exp}(h\xi_k)$ and $\xi_{k+1}=\xi_k+h\,a_\theta(g_k,\xi_k,c_k)$. Here $h$ is the time step and $\operatorname{Exp}:\mathfrak g\to G$ is the Lie exponential. The key property is closure: if $g_k\in G$, then $g_{k+1}\in G$ exactly, unlike an ambient additive update.

Implementation notes

Integrate this at the transition function of a rigid-body world model or robotics policy, specifically replacing the current Euclidean update for rotation matrices, quaternions, or six-degree-of-freedom poses. Use a network $a_\theta$ that receives the current group state $g_k$, Lie-algebra velocity $\xi_k$, and control or context $c_k$, and outputs an element of $\mathfrak g$. For SO(3), represent $g_k=R_k\in\mathbb{R}^{3\times3}$, $\xi_k=\omega_k\in\mathbb{R}^3$, construct the skew matrix $[\omega_k]_\times$, and compute $R_{k+1}=R_k\exp(h[\omega_k]_\times)$. For SE(3), represent the velocity by a six-dimensional twist and use the standard $4\times4$ matrix exponential. Pseudocode: input $(g_0,\xi_0,c_{0:H-1})$; for $k=0,\ldots,H-1$, compute $a=a_\theta(\operatorname{Log}(g_k),\xi_k,c_k)$, set $g_{k+1}=g_k\operatorname{Exp}(h\xi_k)$, then set $\xi_{k+1}=\xi_k+h a$; decode $g_k$ into observations. The paper contributes the forced Euler-Lagrange and trivialized Lie-group structure; the learned acceleration map and discretization are the ML adaptation. Train with geodesic error $d_G(\hat g,g)=\|\operatorname{Log}(\hat g^{-1}g)\|_2$ plus optional acceleration regularization. The first cheap experiment should use a two-layer MLP on synthetic SO(3) rigid-body trajectories, comparing against an equal-size MLP with direct matrix or Euler-angle additions. Measure multi-step geodesic pose error, orthogonality violation $\|R^TR-I\|_F$, and rollout stability for 100 to 1000 steps. Success means zero manifold drift and lower long-horizon error at equal FLOPs.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significant win ✓

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a forced SO(3) dynamics MVP with an MLP acceleration model, exponential-map Lie updates, and an ambient Euler/matrix baseline. The mechanism manifested: Lie orthogonality error was 1.11e-13 versus 4.55 for ambient Euler after 1000 steps; fixed-time ambient orthogonality drift had log-log slope 1.26, consistent with first-order O(h) drift; and constant-velocity ambient geodesic error had slope 1.99 while Lie error stayed around 1e-15. The learned 100-step rollout preserved closure exactly, but did not reduce pose error: mean final-20-step error was 1.397 radians for Lie versus 1.380 for ambient, so no accuracy win is claimed.

Agent confidence: 8/10

Baseline: Ambient Euler: closure violation 4.55 after 1000 steps; fixed-time drift values [4.705, 1.557, 0.638, 0.290, 0.138] for h=[0.2,0.1,0.05,0.025,0.0125]; constant-velocity geodesic errors [0.04698,0.01194,0.002997,0.000750], fitted order 1.99; learned rollout mean final-20 error 1.380 rad, final error 2.345 rad, max orthogonality violation 1.551.
Idea: Lie exponential update: closure violation 1.11e-13 after 1000 steps; constant-velocity geodesic errors approximately [3.83e-16,5.78e-16,7.52e-16,6.02e-16]; learned rollout mean final-20 error 1.397 rad, final error 2.373 rad, max orthogonality violation 3.33e-15. Predicted exact closure and elimination of integration drift were confirmed, but learned pose accuracy was not better.

Limitations: Only SO(3) was tested, not SE(3), rigid-body energy conservation, controls/observations beyond synthetic context, FLOP-normalized training, longer hyperparameter searches, or statistically repeated seeds. The learned comparison is a small toy MLP experiment and evaluates structural stability more reliably than generalization accuracy.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.12627) Sep 2, 2026

Beats the tuned baseline by 75.3% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: So3_forced_rollout
Model: rnn_small
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): -0 (−75.3%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 9 configs

Limitations:

Small registered synthetic SO(3) final-pose regression only; SE(3), real robotics data, energy conservation, longer horizons, FLOP timing, and built-in pendulum transfer were not tested. The completed run used CPU after shared-GPU allocation failures.

How to run: cd /home/maxwelhelp/all/math2nn/experiments/exp820_lie-group-forced-dynamics-layer && OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 CUDA_VISIBLE_DEVICES='' /home/maxwelhelp/main/bin/python3 -u stage2_so3_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Worked ✓

Agent confidence: 9/10

The Lie-group forced dynamics layer significantly outperformed the matched ambient-Euler system on the registered SO(3) dynamics track. Mean test MSE was 4.176081034756862e-06 versus 1.69089952350987e-05, with 8/8 paired wins and permutation p=0.0081. The trained-model signature confirmed near-zero Lie orthogonality error versus ambient drift.

Baseline: Ambient Euler, tuned lr=0.006 step=0.08: mean MSE 1.69089952350987e-05.
Idea: Lie exponential, lr=0.006 step=0.08: mean MSE 4.176081034756862e-06; paired delta -1.2732914200341838e-05; p=0.0081.

Benchmark result