When a robot policy is pretrained on many tasks and then taught a new one from a few dozen demonstrations, more pretraining can make things worse. The LIBERO benchmark found exactly this: imitation pretraining on 90 tasks hurt later lifelong learning, and naive sequential fine-tuning beat methods designed to protect old skills. This proposal borrows a measurement from the physics of glasses. Physicists call it the two-time correlation: how much a system still resembles itself a time $t$ after it reached age $t_w$. We treat a policy's internal features as a system that ages during training. The hypothesis is that an older policy takes proportionally longer to adapt, so the cost of learning a new task from scarce data can be predicted from how long the policy was trained. That prediction is cheap to test and could easily fail.
In glasses and spin glasses, "aging" means relaxation keeps slowing the longer the system has existed. A sample aged ten times longer needs about ten times longer to change by the same relative amount. If robot policies behave like this, three practical consequences follow. A small lab can measure a policy's age directly from its checkpoints. It can predict how many fine-tuning steps a new few-shot task will need. And it can decide when to deliberately "rejuvenate" the policy, a term for briefly perturbing the system so it becomes adaptable again, before spending scarce demonstrations on it. The physics is used only as a measurement tool and a source of falsifiable predictions. We do not claim that SGD is a spin glass.
Core Claim
Let $\phi_s(x)$ be the feature vector a policy computes for observation $x$ at training step $s$. We fix a probe set $P$ of observations from held-out tasks and define the two-time correlation of the policy as
$$C(t_w+t,\,t_w)=\frac{1}{|P|}\sum_{x\in P}\cos\big(\phi_{t_w}(x),\,\phi_{t_w+t}(x)\big),$$
where $t_w$ is the checkpoint's age in steps and $t$ is the further training interval. The claim has two parts.
The first part concerns pretraining. During multi-task pretraining, $C$ stops depending on $t$ alone and instead collapses onto a function of $t/t_w$. This is the "simple aging" form that Phase Space Diffusion and Low Temperature Aging derives exactly, $C\simeq t_w/(t_w+t)$, for a system that slows down because better states become exponentially rare rather than because barriers grow.
The second part concerns adaptation. When a checkpoint of age $t_w$ is fine-tuned on a new task with about 50 demonstrations, the number of steps $\tau$ needed to reach a fixed success rate scales as $\tau\propto t_w^{\mu}$, with $\mu$ near the exponent fitted from the collapse of $C$. If $\mu\approx 1$, the cost of adapting is set by the policy's age and not by the content of its pretraining data.
The novelty is a quantitative, checkpoint-level account of why pretraining hurts few-shot adaptation in robot imitation, together with a test of whether that account predicts transfer better than step count does. Neither LIBERO nor the physics papers attempt this.
Relation to Prior Work
Several papers motivate the question. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning supplies the empirical puzzle: pretraining on LIBERO-90 hurt downstream lifelong learning, and Experience Replay and EWC transferred forward worse than sequential fine-tuning. It also reports a large architecture effect, with forward transfer of 0.70 for ViT-T against 0.30 for ResNet-RNN on LIBERO-Object. LIBERO offers no mechanism for either finding. We test whether one quantity, measured age, accounts for both. Ergodicity in reinforcement learning argues that non-stationary learning breaks the equivalence between averages over time and averages over many runs. It explicitly notes that no principled metric exists for how non-ergodic a learning process is. Our two-time correlation is a candidate for that metric, applied to the learner and not to the environment. That paper tests only low-dimensional tasks such as Cart-pole and Reacher. We work with visuomotor manipulation policies.
The method comes from three physics papers. Phase Space Diffusion and Low Temperature Aging supplies the $t/t_w$ scaling form and the entropic interpretation. That interpretation fits gradient descent better than a story about growing barriers, because a well-trained network sits where directions that still lower the loss are rare. Universal activated aging and weak ergodicity breaking in spin and structural glasses shows that deviations from simple aging, such as Gaussian corrections to trap depths, carry information. We reuse this idea: if collapse holds only with a sub-linear exponent, the fitted exponent becomes the quantity of interest. Quantifying Memory in Spin Glasses shows that a temperature cycle can reset a spin glass's apparent age and that returning to the original temperature recovers the aged state. We borrow the protocol shape, perturb then return, as an intervention. We do not borrow the claim of temperature chaos, which has no established analogue in SGD.
One constraint comes from the physics itself. The 1D memory-as-disorder result in Diffusion in disordered media as a process with memory is rigorous only in one dimension, and the Barrat–Mézard model assumes full connectivity with no correlations between states. We therefore treat every physics prediction as an empirical hypothesis about networks and never as a derivation.
Research Program
The program has three stages.
The first stage measures aging. We train the LIBERO policy architectures (ResNet-RNN, ResNet-T, ViT-T) on LIBERO-90 and save checkpoints at geometrically spaced ages. We then compute $C(t_w+t,t_w)$ on probes drawn from LIBERO-Long, LIBERO-Object and LIBERO-Goal. Each architecture is also run with a diffusion action head in the style of Diffusion Policy: Visuomotor Policy Learning via Action Diffusion, to check that the effect is not tied to one way of outputting actions. We test collapse against $t/t_w$, against $t$ alone (no aging), and against $t/t_w^{\mu}$ (sub-aging). We hold the learning-rate schedule constant after a warm-up. Otherwise a decaying learning rate would trivially produce apparent aging, and that confound is the first thing a reviewer will raise.
The second stage links age to adaptation. Each checkpoint is fine-tuned on each held-out task using LIBERO's 50 human demonstrations, at a fixed budget. We record the full success-versus-step curve and extract $\tau$. The key test is whether curves from different ages superimpose when the step axis is divided by $t_w^{\mu}$. A secondary test asks whether architecture-level differences in fitted aging rate line up with LIBERO's reported forward-transfer gap. Concretely, does ViT-T age more slowly than ResNet-RNN?
The third stage is an age-triggered rejuvenation cycle. When measured $C$ over a sliding window falls below a threshold that signals "frozen," we apply a short perturbation: an elevated learning rate plus small multiplicative weight noise. Training then returns to the base schedule. We measure two things. Rejuvenation is whether new-task $\tau$ drops. Memory is whether old-task success, scored as LIBERO's negative backward transfer, recovers after the return. The intervention is deliberately simple. What is new is that it is scheduled from a measured state of the policy rather than a fixed calendar, and the experiments must show that this scheduling matters.
Baselines
The first baseline is the naive reference: fine-tune from the final pretrained checkpoint, as is standard. In addition, train from scratch on the new task's 50 demonstrations. Comparing the two reproduces LIBERO's finding that pretraining can hurt, in our own runs, before we explain it. Both use identical data and fine-tuning budgets.
The strongest existing lifelong method is sequential fine-tuning, which LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning found often beats ER and EWC on forward transfer. On architecture, the best reported forward-transfer figure in the file is ViT-T at 0.70 on LIBERO-Object. For forgetting, PackNet's reported negative backward transfer is 0.06 on LIBERO-Goal and 0.07 on LIBERO-Spatial. We rerun SEQL, ER, EWC and PackNet with LIBERO's protocol and give each the same number of hyperparameter trials as our method. The rejuvenation cycle beats them if it raises forward transfer above SEQL on the same suite while keeping negative backward transfer within PackNet's reported range. Matching only one of the two is not a win.
The baseline most likely to threaten the intervention is fixed-schedule learning-rate restarts and shrink-and-perturb weight noise. Each gets an equal tuning budget and the same total perturbation magnitude. The file contains no reported result on Lacuna for warm restarts or shrink-and-perturb on LIBERO, so we establish these numbers ourselves.
The diagnostic has its own baselines: predicting $\tau$ from raw step count, from training loss, and from weight-norm growth. The aging metric beats these only if it explains more variance in $\tau$ across architectures and tasks than every simple proxy does. Matching step count would mean the physics adds nothing.
Two ablations isolate the key ingredient. One keeps the same perturbation but triggers it at random times. The other uses a constant learning rate with no cycle.
Why not just…?
Why not just skip pretraining and use sequential fine-tuning, since LIBERO says it already wins? Skipping pretraining throws away what LIBERO-90 teaches about perception and motor primitives. That is exactly what a lab with scarce data cannot afford. The first two stages tell us whether some intermediate age gives the best trade-off. If the success-versus-age curve is monotonically decreasing, the reviewer is right, and we report that pretraining in this setting should be avoided entirely.
Why not just restart the learning rate before each new task? This is the real threat to stage three. If fixed restarts match age-triggered cycles, the intervention is not new, and the contribution shrinks to the diagnostic plus the scaling law. The fixed-restart and random-trigger baselines settle this directly. We commit in advance to reporting it either way.
Why not just generate more demonstrations for the new task with MimicGen-style synthesis? RoboCasa: Large-Scale Simulation of Household Tasks for Generalist Robots shows that synthetic data helps atomic tasks, raising success from 28.8% to 47.6% with 3,000 generated demonstrations. Composite tasks still stay below 12%, and synthesis needs a simulator that many real-robot settings lack. More importantly, adding data and reducing age are complementary. We will run one condition on RoboCasa atomic tasks, crossing checkpoint age with synthetic data volume, to see whether age effects persist when data is plentiful.
Why not just use an existing plasticity measure such as feature rank or the fraction of inactive units? The file contains no reported result on Lacuna for these measures on robot policies. We include them in the stage-two prediction comparison. The two-time correlation is distinctive because it is a relational quantity between two checkpoints, which makes a scaling collapse possible. A single-checkpoint statistic cannot show a collapse.
Decisive Checks
The program succeeds or fails on three outcomes, in order.
The first is collapse: $C(t_w+t,t_w)$ either superimposes as a function of $t/t_w^{\mu}$ across at least a decade of ages under a constant learning rate, or it does not. If it does not, the aging hypothesis is wrong for these networks.
The second is predictive power: on held-out LIBERO tasks, the fitted age predicts adaptation cost $\tau$ better than step count and loss. If it does not, the diagnostic is redundant.
The third is the scheduled intervention: age-triggered rejuvenation beats fixed restarts and random triggers on forward transfer while matching PackNet-level forgetting. If it does not, we keep the measurement and drop the intervention.
Each outcome is informative on its own. If all three hold, small robotics labs gain a principled rule for deciding when a pretrained policy is too old to spend scarce demonstrations on.
References
- Phase Space Diffusion and Low Temperature Aging
- LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning
- Ergodicity in reinforcement learning
- Universal activated aging and weak ergodicity breaking in spin and structural glasses
- Quantifying Memory in Spin Glasses
- Diffusion in disordered media as a process with memory
- Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
- RoboCasa: Large-Scale Simulation of Household Tasks for Generalist Robots