Action generation stops stitching chunks: ODEAction evolves actions by ODE and lifts EBench average success rate by 61.8% relative
Synopsis
Vision-language-action models usually predict a whole future action chunk in one pass, and adjacent chunks lack explicit temporal consistency constraints, so actions can jump at chunk boundaries. ODEAction recasts generation as continuous evolution from the current action: it learns an action transition field and integrates it with an ordinary differential equation to produce subsequent actions, instead of jointly predicting a fixed-length chunk. Average success rate rises from 96.85% to 98.25% on LIBERO and from 15.28% to 24.72% on EBench, a 61.8% relative gain; on five RoboTwin tasks it reaches 84.4% average success against the 82.4% baseline.
Interpretation
The unit of action generation changes from a whole future action chunk to the continuous evolution of the current action, so adjacent actions gain an explicit generative dependency. Earlier methods still treat action chunks as the generation unit: RTC constrains new chunks at inference, and Legato and related work add cross-chunk continuity constraints during training, improving the joins between chunks rather than the generation paradigm itself. Average success rate rises from 96.85% to 98.25% on LIBERO and from 15.28% to 24.72% on EBench, the latter a 61.8% relative improvement, with success rate and all four trajectory metrics improving together across every evaluation unit of both benchmarks.
The method learns an action transition field that, given the current action, an evolution time and visual, language and proprioceptive conditioning, outputs the current action's rate of change with respect to evolution time. Training interpolates discrete demonstrations into a continuous trajectory and samples query actions in its neighborhood; the target field combines a term advancing along the demonstration with a correction term pulling deviating actions back, and is supervised directly without backpropagating through numerical integration. Inference integrates the field step by step with forward Euler, each generated action seeding the next; on all 50 RoboTwin tasks with StarVLA-OFT as the base policy, average success rate rises from 85.7%/87.3% to 89.4%/89.8% under the Clean/Random settings.
Removing neighborhood sampling or the correction term degrades recursive generation sharply, so the gains do not come from dropping action chunks alone. Supervision only on the demonstration trajectory gives 47.4% average success, keeping the neighborhood but dropping correction gives 56.2%, and the full method gives 84.4%; a direct next-action predictor reaches 78.4%. On the same five RoboTwin tasks, offline rollout error is 4.25 for the full method versus 8.84 and 13.17 for the two ablated variants, with error accumulation ratios of 0.92, 2.06 and 2.19.
Perspective
Evaluation covers three simulation benchmarks, LIBERO, RoboTwin and EBench, plus four tasks on an AgileX Cobot Magic 1.0 real robot, with 100 demonstration trajectories per real task and 20 trials per method; within each comparison group the vision-language backbone, training data, control frequency and observation-update times are held identical.
The strong positive correlation between reduced boundary discontinuity and success-rate gains comes from a rank correlation over 10 evaluation units and is a statistical association, not a causal conclusion; real-robot results rest on 20 trials per task, and trajectory metrics are computed only on each method's successful episodes.