Robot demonstrations may contain useful behavior even when individual episodes are inefficient or unsuccessful. Trajectory stitching offers a way to compose these behaviors into improved training data, but identifying useful connections and verifying their feasibility is difficult in high-dimensional robot data, where many prior methods rely on low-dimensional state representations. We introduce NEEDLE, an offline dataset-augmentation algorithm that addresses these challenges by adding short, verified action bridges between recorded observations in high-dimensional robot demonstrations. First, NEEDLE identifies and creates connections that bypass suboptimal detours, broaden action coverage, and augment the original dataset with failed trajectories, using only RGB images, proprioception, and episode-level outcomes, without new environment interaction or privileged object state. Next, we present a sampling technique that incorporates accepted bridges into policy training without synthesizing intermediate images or discarding the original demonstrations, allowing policies to learn alternative actions while retaining the original dataset's coverage. On real-robot tasks, NEEDLE improves success rate over the strongest baseline on each task by an average of 21 percentage points. Videos and supplementary materials are on https://needle-work.github.io/.
Figures & tables
Figure 1: Needle produces synthetic bridges within trajectories that bypass inefficient detours, cross-trajectory stitches that discover more successful ways to complete the task, and recover useful behavior from failed episodes.
Figure 2: From candidate connection to verified bridge. Given a source-target pair (s,g) , the inverse dynamics model (IDM) samples K action proposals. The action-conditioned verifier predicts whether and when each proposal reaches g . We select the shortest accepted prefix as a bridge.
Figure 3: Training on a bridge without intermediate observations. Each example pairs a recorded observation with the original actions leading to the bridge (cyan), followed by the generated bridge actions (green). The loss stops when the bridge ends (red). Original demonstration windows remain available as separate training examples (blue).
Figure 4: Real-robot tasks and evaluation. We evaluate Needle on two real-robot bimanual tasks: Sweater Folding and Dish Racking. We show task sequences, reset distributions, success rates, and completion times. Each policy is evaluated 50 times per task at paired starting configurations across all methods. Completion times show mean and standard error over successful trials.
Figure 5: Fewer failures across task stages. Trials completing each successive stage, out of 50 per policy and task, from video review. Endpoint labels give final success counts, and the negative numbers mark each baseline’s largest loss of trials at a single stage.
Figure 6: Policy performance on suboptimal Robomimic demonstrations. (a) Needle achieves higher average success than all baselines. (b) Its successful rollouts take fewer steps on average than DP’s on Can and Square . Bars show the mean and standard error over five training seeds, with 50 held-out initial conditions per seed; completion steps are averaged over successful trials.
Figure 7: Needle finishes more trials within a fixed budget. Labels give final success rates per time limit over the 50 paired trials.
Figure 8: Shorter continuations and higher success within a step budget. For each task, the left panel shows remaining completion steps after executing either the original continuation or an accepted bridge, restricted to pairs where both continuations succeed. On average, bridges lead to shorter remaining continuations. The right panel shows the fraction of matched nominal starts solved within each budget. Needle reaches comparable success sooner and solves more starts by the full horizon. We plot mean and S.E. for five training seeds, each with 50 held-out starts.
Figure 9: Expanded training coverage and recovery from failure states. For each task, the left panel shows successful-demonstration states and starts added by failure funnels, visualized as 90% density regions in a t-SNE embedding of the visual encoder. The right panel measures recovery after resuming from OOD states, where Needle recovers more often than DP and AWR, the non-DP baseline with the highest average success rate across all three tasks on nominal resets. Curves show mean and standard error for five training seeds, each with 50 held-out starts.
Figure 10: Verifier-based selection connects action sensitivity, bridge quality, and policy success. (a) With the source and target fixed, the verifier distinguishes demonstrated actions from logged substitutes on 1,500 held-out pairs per task. (b) Accepted proposals more often land near the target than rejected proposals. Scene gap is the larger of the arm and object position errors, measured only for this diagnostic; Needle itself never observes object positions. (c) Removing verifier filtering reduces downstream success on all three tasks with the IDM and bridge counts per stage held fixed. Bars show success and standard error over five training seeds of 50 evaluations each.
Task
DP
Ours
Coffee
62.0 ±1.2
63.3 ±0.7
Three Piece Assembly
51.3 ±0.7
55.3 ±1.8
Threading
46.0 ±1.2
46.0 ±1.2
Table 1: Policy success with near-optimal DexMimicGen demonstrations. Needle matches or modestly improves on DP across the three tasks. Values are success rates (%), reported as mean and standard error over three training seeds with 50 held-out initial conditions per seed.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 11: Simulation benchmarks. We evaluate dataset rewriting on three Robomimic tasks with suboptimal demonstrations ( Mandlekar et al., 2022 ) and three DexMimicGen tasks with near-optimal demonstrations ( Jiang et al., 2025 ) .
Figure 12: Retraced motion on the real robot. Fraction of gripper travel revisiting a previously exited 2 cm region, over 100 trials per policy (50 per task). Left: pooled over steps. Right: averaged over trials. Bars show the estimate and standard error.
Figure 13: Which connections help the policy? Stage subsets compared with the main-paper DP and Needle references. Bars show mean success and standard error over five training seeds.
Figure 14: How should a policy learn from a bridge? Rows below Needle name alternative ways to construct training examples. Bars show mean success and standard error over five training seeds; DP and Needle are main-paper references.
Figure 15: Do the bridge actions matter? The first row removes every training window that contains bridge actions and keeps the reweighted original windows. Bars show mean success and standard error over five training seeds; DP and Needle are main-paper references.
Task
DP
SBR
AWR
IDQL
SARM
Ours
Coffee
62.0 ±1.2
60.7 ±2.4
52.7 ±1.8
49.3 ±1.8
39.3 ±1.8
63.3 ±0.7
Three Piece Assembly
51.3 ±0.7
44.7 ±1.8
47.3 ±1.3
43.3 ±3.5
26.7 ±1.8
55.3 ±1.8
Threading
46.0 ±1.2
40.7 ±2.4
50.0 ±1.2
40.0 ±1.2
34.0 ±2.0
46.0 ±1.2
Appendix
Table 2: Full baseline comparison on DexMimicGen. Success rates (%) reported as mean and standard error over three training seeds with 50 evaluation initial conditions per seed. Needle has the highest mean success on two tasks; AWR leads on Threading.
Figure 16: Adapting baselines to visual robot control. Hollow bars show released-policy implementations evaluated on cached DINOv3 features and proprioception; filled bars show the adapted and tuned configurations used in Figure 6 . Adaptation improves success across tasks. Error bars show standard errors over three training seeds for original baseline implementations, and five training seeds for our adaptation; zero-height bars indicate measured zero success.
Setting
Value
Policy, IDM, and verifier
Image encoder / resolution
Frozen DINOv3 ViT-B/16 / 224×224
Optimizer / learning rate
AdamW / 10−4
Weight decay
10−6
Adam coefficients / epsilon
(0.95,0.999) / 10−8
Learning-rate schedule / warmup
Cosine decay / 500 steps
Appendix
Table 3: Shared visual and training settings. The encoder and optimizer settings apply to the policy, IDM, and verifier; the remaining rows specify their scope.
Table 5: Verifier architecture and hyperparameters. Sampling and training values describe the Robomimic runs. Shared visual and optimizer settings are in Table 3 .
Table 6: Stitch generation and training-window construction. Shared settings for the nominal Robomimic runs; task-specific candidate budgets appear in Table 7 . Weights control how often a window is sampled, rather than scaling its action loss.
Setting
Can
Square
Transport
Observations, data, and evaluation
RGB cameras
External + wrist
External + wrist
2 shoulder + 2 wrist
Action dimensions
10
10
20
Successful / unsuccessful episodes
40 / 150
40 / 150
40 / 150
Full evaluation budget
500
500
700
Evaluation starts per seed
50
50
50
Appendix
Table 7: Robomimic task settings. Columns describe the nominal policies used for the main success comparison. Shared training settings appear in Table 3 .
Robot manipulation policies trained through imitation learning commonly rely on costly demonstrations recorded in one task direction and at one temporal resolution. We propose Reversal and Stride-based Trajectory Augmentation (RESTA), which exploits time-scaling data augmentation to improve data efficiency through trajectory reversal and temporal resampling. Trajectory-reversal augmentation reverses and relabels demonstrations between paired tasks with opposite objectives, reusing recorded rollouts without additional environment interaction. Because contact-rich manipulation is not exactly reversible, we treat reversal as an approximate paired-task prior rather than a physical symmetry. Temporal-stride augmentation varies observation and action sampling strides and conditions the policy on the resulting offsets, enabling fixed or randomized temporal schedules. Experiments across 16 paired bimanual manipulation tasks show that trajectory reversal is particularly effective in the low-data regime and that temporal-stride augmentation improves success under randomized temporal schedules. With 25 and 50 independently collected demonstrations per task, RESTA achieves 78.4% and 85.2% average success, respectively, matching or exceeding baselines trained with twice as many demonstrations (77.3% and 85.4%). These results show that RESTA can improve data efficiency in robot imitation learning.
Yunpeng Jiang, Jianshu Hu, Han Fang +3
Global College, Shanghai Jiao Tong University, Shanghai, China · Shanghai Institute of Physical Intelligence and Robotics (PAIR), Shanghai, China · Duke Kunshan University, Jiangsu, China
Adapting visuomotor policies to new manipulation tasks often requires substantial manual engineering or teleoperated data collection. Simulation can provide task-specific data at scale, but constructing the scene, designing expert behavior, and configuring data generation still require significant per-task effort. We present ARSTAG, an agentic Real2Sim2Real system that turns a single RGB image and a natural-language instruction directly into robot policy-learning data. A hierarchy of language agents constructs a task-scoped simulation scene, generates robot-feasible demonstrations, and expands the training distribution through task-consistent randomization, while a coordinator agent manages cross-stage feedback and recovery. Across seven manipulation tasks spanning grasping, placement, and stacking, the ARSTAG-generated demonstrations enable sim-to-real transfer of three visuomotor policy architectures to a dual-arm robot, with pi0.5 achieving an average real-world success rate of 74.6%. Ablations show that task-consistent randomization substantially improves robustness, and policy performance increases with generated dataset size. Project webpage: https://boweili666.github.io/ARSTAG/.
Bowei Li, Yuner Zhang, Changliu Liu
Carnegie Mellon University, Pittsburgh, PA 15213, USA
Robots can acquire bad habits from a few defective moments in otherwise useful teleoperation. In mixed-quality robot data, normal and problematic demonstrations share most task behavior and differ only at a local action continuation. Full retraining is costly, while fine-tuning on clean data alone offers limited recovery under a small update budget. We ask whether the local difference can instead be fixed with only 1% of the full-training sample-backward budget. Using only episode-level retained/problematic labels, REDIRECT localizes the branch, assigns a coherent retained continuation to the original observations around it, and anchors shared behavior, without frame-level annotations or additional interaction. Across three randomized ManiSkill tasks and three seeds, REDIRECT raises the mean success rate from 68.2% to 90.3%, recovering 88.8% of the clean-retraining gap versus 22.1% for matched-compute fine-tuning. On a PiPER arm, it recovers 86.7-92.9% of the clean-only gap across cup insertion and towel folding. Local robot habits can therefore be repaired by spending the update budget on the behavioral difference rather than relearning shared behavior.