Robot controllers trained in simulation are usually made robust by domain randomization: during training, the simulator's physical parameters (masses, link lengths, friction, motor strength, delays) are drawn at random each episode, so the controller learns to cope with a whole family of slightly wrong worlds. This works only if the real robot falls inside that family. When the real payload is heavier, or the link longer, than anything sampled, neither theory nor practice offers much. Physics has long had a partial answer to this kind of extrapolation. Two systems that differ in size, mass and speed behave identically once their variables are measured in each system's own natural units. A pendulum ten meters long is a slowed-down, scaled-up copy of a one-meter pendulum. The quantities that decide whether two systems are copies are the dimensionless groups of the Buckingham π theorem (unit-free ratios such as torque limit over weight times length).
We propose to train and run robot controllers entirely in these dimensionless coordinates. Observations, actions and time are rescaled by characteristic scales computed from the robot's known nominal mass, length and gravity. Randomization is then applied to the dimensionless groups rather than to raw masses and lengths. The consequence is concrete and testable. A real robot whose raw parameters lie far outside the training range, but whose dimensionless groups lie inside it, becomes an in-distribution problem for the controller with no retraining. The proposal makes no promise for shifts that change the dimensionless groups themselves, and the experiments are built to show that boundary explicitly.
Core Claim
The standard theory of domain randomization, Understanding Domain Randomization for Sim-to-real Transfer, models the simulator as a hidden-parameter decision process. It bounds the sim-to-real gap only when the real parameter $\theta^*$ lies in the support of the randomization distribution, and is silent otherwise. That analysis motivates this work by locating the gap precisely. Sim-to-Real Transfer of Robotic Control with Dynamics Randomization states the same gap empirically as an open question. Our claim is that the support condition is being checked in the wrong coordinates. Let $\Pi(\theta)$ denote the vector of dimensionless groups of parameter vector $\theta$. If the controller observes, acts and keeps time in units fixed by $\theta$'s own characteristic scales, then exact dynamic similarity implies that the theory's guarantee needs only
$$\Pi(\theta^*) \in \mathrm{supp}(\Pi_\# \mu),$$
where $\mu$ is the training distribution over parameters and $\Pi_\# \mu$ is the distribution of dimensionless groups it induces. This is a strictly weaker requirement than $\theta^* \in \mathrm{supp}(\mu)$. It holds because similar systems produce identical trajectories in dimensionless coordinates, so they are the same hidden-parameter instance. The claim is new in two ways. It moves the Buckingham π idea from static field prediction to closed-loop control. It also gives a sharp, falsifiable prediction about which out-of-range robots a randomized controller can and cannot handle.
Research Program
The construction is simple to state. For an articulated robot with nominal characteristic length $L$, mass $m$ and gravity $g$, define a characteristic time $\tau = \sqrt{L/g}$ and torque scale $mgL$. The controller then sees two kinds of quantities. Joint angles and Coulomb friction coefficients are already dimensionless and pass through unchanged. Velocities are multiplied by $\tau$. Commanded torques are emitted as fractions of $mgL$, and contact forces as fractions of $mg$.
The remaining physics becomes a short list of dimensionless groups that are randomized during training. These are the actuator torque limit over $mgL$, joint damping times $\tau/(mL^2)$, the ratio of each link's mass and length to the nominal ones, the friction coefficient, the control period $\Delta t/\tau$, and the sensing delay over $\tau$.
Two of these deserve emphasis because they are where similarity breaks on real hardware. A physical robot runs its controller at a fixed wall-clock rate and has a fixed bus latency (15–19 ms on the Minitaur in Sim-to-Real: Learning Agile Locomotion For Quadruped Robots). A larger robot therefore sees a smaller dimensionless control period and delay. We treat these as first-class randomized groups rather than pretending they scale.
This rescaling avoids the main obstacles that stopped earlier dimensional-analysis methods from reaching control. BuckiNet in Dimensionally Consistent Learning with Buckingham Pi takes logarithms of its inputs and so cannot accept signed velocities or torques. Buckingham π-Invariant Test-Time Projection for Robust PDE Surrogate Modeling needs global field averages to avoid dividing by zero. In our construction, logarithms are taken only of the positive physical parameters that define the scales, never of the state. Signed states are merely multiplied by positive constants, so no degeneracy arises.
Those two papers supply the method, namely null-space computation of π-groups and the equivalence-class view of scale changes. We depart from them by building invariance into a policy's input–output interface and its training distribution, instead of a test-time projection of a static field. The controller itself remains an ordinary recurrent policy, so the memory-based identification that Chen et al. prove necessary still handles unknown dimensionless groups such as friction.
The analysis component asks how fast performance degrades as the real robot moves off the training set of dimensionless groups. We will also measure how much of a realistic shift is "similarity-reachable", meaning it changes raw parameters while leaving the π-groups in range. Empirical effects such as Stribeck friction, hysteresis and gear backlash have no clean dimensional formula. We will measure how much each of them erodes the guarantee, rather than assume it away.
Baselines
The naive reference is a recurrent policy trained without randomization at nominal parameters, run in raw units. It anchors how badly out-of-range robots fail.
The strongest existing method for this question is standard dynamics randomization with a memory-based policy, in the style of Sim-to-Real Transfer of Robotic Control with Dynamics Randomization. That paper reports a real-robot puck-pushing success rate of 0.89 ± 0.06 for the LSTM with randomization, versus 0.51 for a feedforward policy without randomization. We implement it through the domain-randomization wrappers of RRLS (RRLS : Robust Reinforcement Learning Suite). That benchmark found plain randomization to be the strongest baseline for average-case return, ahead of the adversarial methods it evaluates. The file gives no numerical RRLS returns, so no reported result was found on Lacuna for RRLS out-of-range performance; we will establish it ourselves. Because RRLS reports that M2TD3 is better in static worst-case settings, we also include it as a worst-case-oriented competitor.
A third baseline is a context-conditioned policy that receives the true parameters $\theta$ as extra inputs, in raw units, alongside memory. It is the controller analogue of the context-conditioned world models that Graph-Operator World Models for Morphology-Parameter Generalization in Continuous Control reports failing on extrapolation splits. The file contains no numeric result for that comparison.
All methods share the same simulator, network size, number of environment steps, seeds, and hyperparameter search budget. Randomization ranges are matched so that every method sees the same set of training robots. The only difference is the coordinates in which that set is described and sampled.
Two ablations isolate the key ingredient. The first keeps raw-unit observations and actions but samples the same robots. The second uses dimensionless observations and actions with ordinary raw-unit sampling. Together they separate the effect of the rescaled interface from the effect of the reshaped training distribution.
Beating the strongest baseline means two things. On similarity-reachable out-of-range robots, our policy must retain at least its in-range return or success rate while dimensional randomization degrades, measured across seeds with confidence intervals. On π-breaking shifts, our policy must not be worse than dimensional randomization.
Why not just…?
Why not just widen the randomization range until the real robot is inside it? This is the most natural fix. Dynamics as Prompts: In-Context Learning for Sim-to-Real System Identifications and RRLS both report that wide ranges produce conservative "average" policies with poor nominal performance. Widening is also unbounded: no finite range covers every payload. Our experiment runs a wide-range baseline whose raw-parameter range covers our test robots. If it matches our method at equal compute without losing nominal performance, the core claim fails. We consider this a real threat for mild shifts and expect the separation to appear only at large scale factors.
Why not just use standard observation normalization, as most RL code already does? Running-mean normalization rescales inputs with statistics pooled over all training robots, not with each robot's own scales. It does not rescale actions or time, so a longer link still produces a slower, mis-timed response to the same normalized command. The first ablation above, which keeps normalized raw-unit inputs, directly measures whether this cheaper trick suffices.
Why not just feed the parameters to a recurrent policy and let memory identify the rest? Memory identifies only parameters inside the range it was trained on. CAPTURE's in-context identifier saturates at its search bounds and needs 3 to 4 attempts. Chen et al. prove that memory cannot rescue tasks where one misjudged step is unrecoverable. Our method keeps memory for the unknown dimensionless groups and removes the need to extrapolate in the known ones. The context-conditioned baseline tests this division of labor head on.
Why not just apply the test-time π-projection of Lee et al. to an ordinary trained policy? Projecting a single observation does not fix the action scale or the clock, and closed-loop errors compound. We include this as an extra baseline: dimensional training plus projection of observations at test time. It tells us whether training in π-space adds anything beyond input projection.
Proposed Experiments
The simulation study uses MuJoCo Hopper, Walker2d, HalfCheetah, Ant and InvertedPendulum through RRLS. Parameter splits follow the GraphOp-WM benchmark design: training envelope, interpolation, extrapolation and held-out composition. We add one deliberate twist. The extrapolation split is divided into similarity-reachable robots and π-breaking robots. Similarity-reachable robots scale all lengths by $k$, masses by $k^3$ and torque limits by $k^4$, for $k$ up to 3, which preserves the π-groups. π-breaking robots change one group, for example a weaker actuator relative to weight.
The decisive check is the shape of the results across these splits. Our method should be flat across $k$ on the reachable split and degrade on the π-breaking split at the same rate as dimensional randomization. Dimensional randomization should degrade on both. A result in which our method also degrades with $k$ on reachable robots would falsify the mechanism. To rule out simulator artifacts, we will first verify that such degradation is not caused by integration-step effects; the fixed-delay groups are included for exactly this reason.
A sim-to-sim reality gap comes next, again in MuJoCo Playground for throughput. Policies trained in a clean simulator are tested in one with injected Stribeck friction, backlash and fixed wall-clock latency. This measures how much non-similar physics erodes the reachable-split advantage.
Finally, a zero-shot hardware test uses a physical pendulum swing-up with swappable links and bobs whose length and mass exceed the trained raw range. It runs at a fixed hardware control rate, and the reported metric is swing-up success and time to balance. We also repeat Peng-style tabletop pushing with pucks heavier than any trained mass, reporting success rate against the 0.89 ± 0.06 reference.
The proposal succeeds if the reachable/π-breaking split cleanly predicts where transfer holds on hardware. Even a negative hardware result would be informative if the latency group explains the failure.
References
- Understanding Domain Randomization for Sim-to-real Transfer
- Sim-to-Real Transfer of Robotic Control with Dynamics Randomization
- Sim-to-Real: Learning Agile Locomotion For Quadruped Robots
- Dimensionally consistent learning with Buckingham Pi
- Buckingham $\pi$-Invariant Test‑Time Projection for Robust PDE Surrogate Modeling
- Rrls: Robust reinforcement learning suite
- Graph-Operator World Models for Morphology-Parameter Generalization in Continuous Control
- Dynamics as Prompts: In-Context Learning for Sim-to-Real System Identifications
- MuJoCo Playground