Conditional Trajectory Peaks: Single-Pass Multimodal Policies over Action Chunks

Conditional Trajectory Peaks. Single-Pass Multimodal Policies over Action Chunks. By Di Wu and six others, from Magiclab Robotics and Southeast University. Published on the fifth of October. No code repository is linked to it.

Multimodal imitation learning asks two things of a policy at once. Given the same observation, it should be able to produce diverse futures that can all be executed. And across replanning cycles, its behaviour should stay consistent, not hop between those futures from one cycle to the next. Doing both, and doing it quickly, is the challenge this work addresses.

Conditional Trajectory Peaks, or CTP, is a single-pass policy framework. In one pass it jointly predicts a set of complete action-chunk candidates, a probability mass for each, and a scale for each trajectory. Each candidate is a peak: one possible future, with a weight attached.

Two components shape those peaks. Distribution-Aware Peak Specialization, or DAPS, makes the peaks specialize, using posterior responsibilities computed at the level of whole trajectories, together with overlap constraints modulated by mass and scale. Evidence-Gated Trajectory Belief Transport, or ETBT, handles consistency across chunks. It matches one candidate set to the next through geometric correspondence, treating the candidates as exchangeable, and still lets current policy evidence override the constraints inherited from history.

On Push-T, CTP reaches a coverage score of ninety-one point four percent. On D3IL, its success rates are one hundred percent on Avoiding, seventy-nine point seven two percent on Aligning, and eighty-four point four four percent on Sorting-2. On LIBERO it averages ninety-seven point two five percent.

In real-world dual-arm experiments, CTP keeps both placement modes alive in a two-plate task and succeeds in all fifty trials. On bottle uprighting and on placing a pen into a holder, its success rates are comparable to pi zero point five, while policy inference latency falls from two hundred eighteen point two four milliseconds to seventy-five point eight milliseconds.

The authors' conclusion is that single-pass trajectory modeling can deliver three properties together: multimodal behaviour, consistency in closed loop, and efficient inference. All three rest on one design choice. The policy lays out several weighted futures in a single pass, and then decides, chunk by chunk, how far the past should constrain which of them it carries forward.