STEAM: Self-Supervised Temporal Ensemble Advantage Modeling for Real-World Robot Learning

Reward design remains a central bottleneck for autonomous robot policy improvement, especially in long-horizon manipulation tasks where sparse success labels provide too little signal and binary preferences collapse many competing notions of quality into one ambiguous signal. We introduce Freeform Preference Learning (FPL), a method for learning robot policies from freeform human preferences. Rather than asking annotators which of two trajectories is better overall, FPL lets them define natural-language preference axes, such as speed, safety, quality of placement, or carefulness, and provide pairwise preferences along each axis. These annotations are used to learn a language-conditioned reward model that maps a trajectory and preference label to an axis-specific reward. We use this model to train a reward-conditioned policy that optimizes across the multiple human-specified dimensions. Across four real-world and two simulated long-horizon manipulation tasks, FPL improves over sparse-reward and binary-preference methods by 38 percentage points. Beyond improved performance, FPL learns dense progress signals without explicit subtask segmentation, shows compositionality of behavior not present in the data, and allows users to steer the policy towards different behaviors at test time without retraining. Blog post with videos available at https://freeform-pl.github.io/fpl.website/

Long-horizon robot manipulation policies trained with reward shaping can still achieve high return through inefficient interactions, while rare efficient behaviors discovered during training may be forgotten. We argue that temporal efficiency itself provides a powerful and underutilized source of self-supervision for reinforcement learning. We introduce Temporal Self-Imitation Learning (TSIL), a reinforcement learning framework that mines temporally efficient successful trajectories generated during learning and converts them into reusable supervision for future policy improvement. TSIL progressively refines learning using configuration-conditioned adaptive temporal targets derived from fast successful trajectories, while preserving and replaying efficient behaviors through efficiency-weighted self-imitation learning. Across 15 distinct long-horizon manipulation tasks, TSIL consistently improves learning efficiency, task-completion efficiency, revisitation of fast successful behaviors, and robustness to unstable training conditions. More broadly, our results suggest that the temporal structure of successful behavior itself provides a scalable self-supervisory signal for reinforcement learning beyond manually engineered reward shaping alone.

Expert demonstrations are widely assumed to be the gold standard for robot imitation learning. Yet for fine-grained manipulation such as insertion, stacking, and alignment, we uncover a counterintuitive failure mode: fluent demonstrations can be poor teachers. A skilled teleoperator compresses the decisive moments of alignment and recovery into a brief temporal window, leaving the policy flooded with redundant free-space motion and starved of supervision exactly where precision determines success. We address this bottleneck at two levels. At the data level, slowing down near alignment and resampling critical segments both help, yet the gain comes mainly from broadening the coverage of recovery states the policy must learn, not from reweighting frames it already has. Such data-side fixes, however, leave the policy's per-frame view untouched: a single image still maps directly to an action, and the local motion that governs correction stays implicit. We therefore turn to the representation level and introduce STAIR (\textbf{S}patio-\textbf{T}emporal feature \textbf{A}s an \textbf{I}nterface for \textbf{R}obot learning), a compact dynamic feature that bridges the vision-language model and the action expert, distilling the short-horizon motion already recorded in each trajectory into dense, motion-aware supervision. Trained on fluent data alone, STAIR recovers most of the deliberate-demonstration gain ($50.0$ to $62.2\%$ overall, approaching the $64.4\%$ of deliberate demonstrations). These results call for a more pedagogical view of robot data, optimized for machine learnability rather than human efficiency alone.