LM-X recorded higher average success than comparison systems in three settings: a controlled five-task gate, all 50 tasks in the RoboTwin2.0 benchmark, and seven real-world manipulation tasks. The results are reported in an arXiv preprint, version 5, dated 8 October 2026.
The system is built to expose three directly supervised online signals. RTG represents visible progress and state quality; ETG predicts the next meaningful task event; and a heteroscedastic action-flow variance provides an uncertainty estimate for local command reliability that can change from one action to the next.
Signals built into the controller
LM-X was trained on more than 20,000 hours of heterogeneous real-robot trajectories, including over 1,000 hours of failed policy rollouts. The training material carried outcome, progress and event labels, so unsuccessful policy behavior was part of the dataset.
Event labels were created in two stages. Task-dependent proposals marked likely transitions, then people checked the boundaries and corrected them when they did not match the intended transition.
What the tests found
The controlled gate used the same 45 RoboTwin2.0 pretraining tasks and five disjoint downstream tasks, with a matched optimization budget and evaluation protocol across variants. LM-X reached 79.6% mean success, compared with 63.6% for the action-only backbone, a gap of 16.0 percentage points. It was 10.8 points above the strongest single-component variant and 6.8 points above the 72.8% post-hoc control.
Across all 50 RoboTwin2.0 tasks, the model used 50 demonstrations per task from the randomized-hard split and was tested in 100 trials per task. Its mean success was 74.1%, against 55.4% for GR00T N1.7, an absolute difference of 18.7 points.
The physical evaluation covered seven tasks on four robot embodiments, with 20 trials per task and the same task-specific data and success criteria across methods. LM-X achieved 73.5% mean success, compared with 50.7% for GR00T N1.7 and 36.4% for π0.5. Against GR00T, it did better on five tasks, tied on one and trailed on one.
Numbers that describe the robot’s view
Foundation pretraining was associated with a stronger RTG progress signal. Average progress correlation rose from 0.71 without pretraining to 0.91 with it, while mean squared error, a measure of numerical miss where lower is better, fell from 0.100 to 0.050. For ETG, the average next-event error fell from 0.884 to 0.520, a reported relative reduction of 41.2%, with improvement on all seven tasks.
The uncertainty signal also improved as a warning indicator. Its action-error AUPRC, a score for ranking high-error action chunks, rose from 0.49 to 0.59. Mean gradient-based failure-warning lead time grew from 0.93 seconds to 1.54 seconds, a gain of 0.61 seconds, at a fixed 10% false-alarm rate. Both measures improved on all seven tasks.
Those traces should not be read as direct probabilities that a task will fail. The paper treats RTG and uncertainty as diagnostic predictive signals, not direct failure probabilities.
Where the evidence stops
One gate result points to task-dependent brittleness. The event-only condition succeeded on the open-microwave task 16% of the time, compared with 54% in the comparison condition.
The reported evaluation also has a limited statistical base. The real-world study used 20 trials per task and one checkpoint for each method, with no confidence intervals; the simulation study used a single training run per configuration.
The paper therefore pairs higher reported task success with online predictive signals, while presenting RTG and uncertainty as diagnostics rather than direct failure probabilities. The comparisons remain tied to the tested trials and runs.
Paper data and sources
Original title: LM-X: Explainable Action Modeling with Progress, Event, and Uncertainty Prediction for Generalist Robot Manipulation
Authors: Jin Lou, Jingxuan Zhu, Andong Chen et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text