EqC: Equilibrium Energy Control for Introspective Lifelong Robot Learning
Haodong Yang, Runhan Huang, Bowen Zhang, Iofai Lei, Zunzhe Zhang, Xiangcheng Zhang, Shaoting Zhu, Linzhan Mou, Yicheng Liu, Hang Zhao
cs.RO
2026-10-14
EqC climbs an explicit energy to refine actions and calls an expert when that path breaks the learned relation. Garment folding goes from 30% to 98.3% after 12 recoveries.
Diffusion policies and flow-matching vision-language-action models still go out as static action generators. Images, language, and proprioception come in, an action chunk comes out, and the robot runs a short prefix before replanning. Familiar layouts get finished. Slip, occlusion, or a shifted start pose walks the same policy off the experience that supports it.
The missing piece is a competence check at execution time. A late takeover logs a failure. A trigger-happy takeover cancels episodes the policy could have finished. Lifelong learning needs those calls pressed against the competence boundary. DAgger already specifies rollout, correction, and retraining. The open question is when to speak. Prior gates use predicted action error, ensemble disagreement, novelty, or the flow-matching loss in Diff-DAgger. Terminal energy has been a visual-anomaly score, and had not driven data aggregation on a physical robot. EqC puts the action and the call for help on one conditional energy.
The condition z packs images, language, and proprioception. The output is a continuous action chunk a. Energy is an inner product on the active dimensions, Eθ(x, z)=⟨x, fθ(x, z)⟩. Its gradient is a stationary height map: it depends on the current action and the condition, and it does not change with bridge time. Equilibrium Matching and GeCO leave energy implicit in the dynamics. Once the scalar is explicit, the climb has a checkable energy budget, and the action update is still that same gradient.
Training follows a noise-to-expert bridge. t=0 is noise, t=1 is data, d=a−ε, and xt=ε+td. Gradient matching aligns ∇x E with c(t)d. The reported schedule is c(t)=min{4, 20(1−t)}. Time shows up only in the training target. Inference starts from Gaussian noise, ascends the energy, executes a short prefix, and replans. Stable steps make the energy rise monotonically. Diffusion Transformers and the flow-matching VLAs π0 and π0.5 share the interface. On the VLA, the multimodal prefix and the action expert stay; model time is fixed at 0, and E=⟨x, vθ(x, o, 0)⟩ replaces flow matching.
Introspection reads the whole path, not the size of the terminal energy. If gradient matching holds, the energy gap out to the expert action equals ‖d‖² times the remaining schedule integral. Deployment has no expert action, so the optimization endpoint is the anchor. Each iterate is projected onto the chord from noise to that endpoint, and UΔE is the mean absolute gap between normalized energy-to-go and the bridge reference. A path that follows the learned geometry scores low. No extra uncertainty model is trained.
Energy-Gated DAgger reads UΔE before every prefix. Only m consecutive crossings of γ hand control to the expert, and the recovery enters the next round. The physical loop uses γ=0.2. A single spike does not fire.
In-distribution success moves by about a point. Perturbation, failure detection, and the physical loop are where the gap opens.
| Setting | Baseline | EqC |
| LIBERO-DiT mean | DP 86.5%, RF 90.0%, GeCO 93.5% | 94.0% (multi-seed 93.95±0.66) |
| LIBERO, π0 | 94.2%, GeCO 93.9% | 94.3% |
| LIBERO-Plus | π0 53.6% | 65.8% |
| RoboTwin easy/hard | π0 49/18, GeCO 50/28 | 63/33 |
| VLABench | π0 29.4%, π0.5 24.5% | 36.7%, 27.1% |
| Real robot, 20 trials each | G0.5 63.3% | 91.7% |
| Failure-detection AUROC | Diff-DAgger 0.8430 | 0.9244 |
| Block stacking, 30 recoveries | start 0%, Diff-DAgger 80.0%, human 58.3% | 96.7% |
| Garment folding, 12 recoveries | start 30.0%, Diff-DAgger 81.7%, human 55.0% | 98.3% |
On LIBERO-Plus, camera viewpoint goes from 15.8% to 61.7% and robot initialization from 6.6% to 39.0%. Language falls from 61.0% to 54.8%, and sensor noise from 79.4% to 74.3%. On the DiT Object suite, published GeCO at 99.0% still beats EqC at 98.1%. Long is 85.5% for EqC, 76.0% for rectified flow, and 83.8% for GeCO. A matched multi-seed rerun reports rectified flow at 89.12% and GeCO at 92.30%. Under RoboTwin Hard, Adjust Bottle moves from 56% to 76% and Click Alarmclock from 11% to 32%. Size ranking stays near zero. VLABench T4 drops from 17.3% to 15.8%.
The static real-robot trio is a separate experiment: flower insertion 45% to 80%, towel folding 80% to 95%, debris sweeping 65% to 100%. Mean replan latency rises from 432.4 ms to 543.9 ms. Garment folding and block stacking above are the lifelong runs, not this trio. Failure labels are final outcomes. A 10% split freezes the window and threshold with false positives held near 10%. At the reported point m=11, true-positive / false-positive rates go from 61.3%/10.2% to 77.9%/8.5%, and time saved on failed rollouts goes from 33.3% to 45.5%. The flow-loss detector on the EqC backbone reaches AUROC 0.8888, above the original Diff-DAgger and below the path score. Terminal energy scores only 0.52 to 0.65. The full path reaches 0.930 at m=13. Raising the Monte Carlo count to 128 does not close the gap. On physical block stacking, 25 successes and 25 failures separate early under UΔE, while the Diff-DAgger curves overlap for most of the horizon. Lifelong numbers average three runs, with 20 autonomous trials per checkpoint per run. After a later shift into cup stacking, the main text reports dual-task autonomy back at 100%.
A DiT or π0 stack can keep its encoders and action interface. What changes is the action objective and an ascent along the energy at inference. One energy both commands the robot and decides whether autonomy should continue. The expert action is not required at deployment.
Twelve garment recoveries and thirty block recoveries turn assisted situations into near-complete autonomy. A human pressing the takeover button on the same budget stays around 55% to 58%. The one-point in-distribution gains are incremental. The useful part is robustness under camera and initialization shift, and a gate that spends expert time on the boundary. The expert is still in the loop.
The study stays with continuous action chunks and a stationary refinement schedule. Hybrid, memory, and hierarchical policies are untouched. The paper states that thresholds and intervention budgets are untested across embodiments and longer deployments.
Averages hide regressions. Language shift, sensor noise, and VLABench T4 are worse than π0. Size ranking is near zero for every method. The energy geometry stabilizes the optimization path. It does not add semantic understanding.
The lifelong curves rest on three runs and 20 trials per checkpoint, with no per-round standard deviation in the main text. Debris sweeping at 100% is 20 out of 20. The human gate trails by a wide margin. That can be late timing, or a difference in how long and how clean the teleoperated snippets are. The paper only matches data volume and training budget. γ=0.2 and the persistence window are calibrated, and the main text does not say whether a new task needs a new calibration. The 100% dual-task claim is one sentence, without a forgetting curve. When actions are clearly multi-modal, the approximation cost of a stationary energy against a time-conditioned field is also unmeasured. Published GeCO on LIBERO-DiT is 93.5%; the authors' own multi-seed rerun is 92.30%. That gap belongs in any reading of the headline table.