AdaTempo: Learning Shared Relative Tempo from Demonstrations for Faster Robot Manipulation
This preprint presents AdaTempo, a self-supervised method that speeds up visuomotor imitation policies by extracting shared phase-relative tempo patterns from demonstrations and resampling trajectories before training. It achieves up to a 3.57x speedup over demonstrations, and in real-robot tests raises aggregate success from 58.3% with ACT to 77.8% while completing tasks 2.02x faster. The approach matters because it embeds faster execution directly into standard policies such as ACT or Diffusion Policy without runtime retiming or extra speed-selection modules.
Can demonstrations themselves provide a direct and fine-grained signal for phase-dependent speedup, rather than relying on auxiliary runtime mechanisms or policy-derived proxies?
Visuomotor policies trained by imitation inherit the slow, cautious timing of teleoperated demonstrations, but uniform speedup is unreliable because different task phases tolerate acceleration differently. Existing acceleration methods typically rely on auxiliary models, environment interaction, or runtime speed-selection machinery, adding complexity and often reducing task success.
Prior families include runtime execution acceleration such as SAIL, SpeedTuning, and variable-frequency conditioning, as well as demonstration-side methods like DemoSpeedup, AutoSpeed, SpeedAug, ESPADA, and fixed-rate downsampling baselines such as ACT-2x or DP-Fast. These approaches generally require policy-derived entropy or error proxies, reinforcement-learning interaction, semantic labels, or runtime speed scheduling, and they often trade away success for speed.
AdaTempo phase-aligns demonstrations using dynamic time warping with a motion-progress regularizer, converts each episode into relative tempo by ranking inverse phase occupancy, and aggregates cross-demonstration tempo consensus with confidence weighting based on median absolute deviation. It maps this consensus to a continuous speedup profile using lookahead soft-minimum and asymmetric slew limiting, then resamples source trajectories by varying virtual time. The accelerated trajectories are used to train standard ACT or Diffusion Policy backbones, with no runtime tempo-selection module.
Across five simulation tasks, AdaTempo averaged 75.5% success at a 2.83x speedup over demonstrations, versus DemoSpeedup at 68.1% success and 2.86x SOD. On ALOHA Transfer Cube, ACT+AdaTempo reached 83.67% success versus 45.33% for ACT, 53.00% for ACT-2x, 69.00% for ACT+DemoSpeedup, and 63.33% for SAIL. On Insertion, success was 25.00% with AdaTempo versus 23.33% for ACT. In real-world tests across six tasks, AdaTempo achieved 77.8% aggregate success versus 58.3% for ACT and 56.7% for DemoSpeedup, with mean task completion 2.02x faster than ACT. On Conveyor Fast, it succeeded in 29/30 trials at 3.2s versus 2/30 for ACT and 12/30 for DemoSpeedup. Cross-demonstrator tempo profiles on Conveyor Fast had pairwise Spearman correlations of 0.983, 0.941, and 0.955; DTW alignment improved leave-one-out tempo agreement to 0.861-0.952 across tasks versus 0.662-0.885 with arc-length-only correspondence.
The authors state that relative tempo is empirical rather than a direct estimate of feasible speed, shared teleoperation habits may bias it, and a single medoid/consensus may fail with unrelated pauses or qualitatively different strategies. Acceleration bounds are task-specific and selected using validation rollouts. Resampling cannot remove physical limits, tracking error, actuator limits, latency, or contact dynamics, and high-gain control adds no formal safety guarantee. The learned tempo is fixed at deployment and cannot adapt online. Evaluation is limited to one real bimanual ALOHA platform with 30 trials per task and five simulation tasks, and the paper does not report code release or computational cost.
Robot integrators and companies deploying bimanual manipulation, warehouse picking, assembly, or food-handling systems could use AdaTempo with existing ACT or Diffusion Policy pipelines to reduce cycle times without deploying extra runtime controllers. Near-term impact in 1-3 years is plausible if automatic acceleration-bound selection, broader robot/hardware validation, and safety checks are added; adoption on production hardware also requires confirming tracking quality at higher speeds and handling multi-strategy demonstration data.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.