NEEDLEwork: Rebuttal Materials

1. What if bridges don't match the remaining actions?

Reviewers 4XGn and voED

Our training method does not assume an exact connection between a bridge and the target's remaining actions. We see that the verifier plays an important role in filtering out harmful errors, and our training method ensures we're able to learn well in spite of any remaining imprecision. Training windows stop action supervision at the bridge endpoint (Figure 3 of the main paper). The target's remaining actions are learned separately, paired with their own recorded observations. Appending those actions directly to the bridge reduces downstream success on all three tasks. We also test whether a closed-loop policy can continue from bridge landings.

How NEEDLEwork builds training windows from a bridge (animation)
Each window pairs a recorded observation with the actions leading into the bridge (cyan), then the bridge actions (green). Sliding the start exposes the connection from several preceding observations. Slots after the bridge carry no training loss. The target continuation (blue) and the original source and skipped-segment actions remain separate training examples. Animated on Figure 3 of the main paper.
The append ablation in Figure 1.1. Where a window runs past the end of the bridge, NEEDLEwork leaves the remaining slots unscored. The ablation fills them with the target demonstration's next actions instead, which treats the target's recorded continuation as valid from the bridge landing.

Downstream success under different training designs ↑ Higher Is Bettermean ± standard error over five seeds

Figure 1.1: Appending the target's remaining actions gives lower mean success than NEEDLEwork on all three tasks. Removing the verifier gives means near or below DP, with the largest drop on Transport. Values are per-seed best evaluation success, averaged over five training seeds; Appendix B.3 reports the other training-window ablations.

This suggests that bridges do not need to perfectly connect to the target's suffix trajectory to be usable for policy learning.

Success from the bridge landing, relative to the original dataset → Higher Is Better

Figure 1.2: For each bridge in the training data, we compare the success of open loop replay of target trajectory's actions (grey), and closed loop execution of a diffusion policy (blue), against the original dataset's actions. A diffusion policy that takes over from the same landings succeeds as often as it does after the source's own recorded actions.
Back to contents ↑

2. Which bridge errors hurt, and does the verifier remove them?

Reviewers 4XGn and voED

We analyze two different types of bridge errors that occur in our dataset: pose error, where the arm lands >5cm from the target location, and grasp error, where the gripper drops the object during the bridge. We find that grasp errors hurt the policy more, lowering its performance below DP. Our verifier keeps far fewer of these errors than a value function.

Success after training on one type of bridge error ↑ Higher Is Bettermean ± standard error over three seeds

Figure 2.1: To test which error type matters, we keep the paper's Can, Square, and Transport training sets (blue) fixed in size and composition (same number of bridges per stage, same source states where possible) and swap their bridges for bridges of one error type: pose errors (light grey) or grasp errors (dark grey). Each replaced bridge is paired with one of the same stage, similar length, and similar steps saved, so the two sets differ mainly in error type. Grasp errors cost more than pose errors on all three tasks (Can −4.7 ± 0.7, Square −6.7 ± 1.8, Transport −8.7 ± 5.8 percentage points). All three bars use the same three training seeds. Dotted lines: DP on the same three seeds (72.0 / 40.0 / 47.3).

How much of the training data carries each error

Figure 2.2: Top bar of each task: the share of training windows that contain a bridge action. Bottom bar: the bridges kept by our verifier (the same bridges as in Figure 2.3), split by error type. Click on each segment to see an example.

Grasp errors among kept candidates ↓ Lower Is Better

Gripper landing error of kept candidates ↓ Lower Is Better

Figure 2.3: Left: We compare the amount of grasp errors falsely admitted by our verifier, compared to the value function learned and used by IDQL and AWR. Right: We calculate the average pose error of the end effector with the same selection criterion used by the value function above. Our verifier admits fewer false positives on both error types, which helps explain why stitching with our method performs better than the value-function alternative. Click on the value function and rejected bars to see examples.

3. Impact of quality and quantity of data on stitching.

Reviewer ak1N

Stitching helps more as the quality of data decreases. Over different quantities of total data, stitching's gain is preserved.

Success rate under different mixtures of suboptimal data ↑ Higher Is Bettermean ± standard error over five seeds

Figure 3.1: We vary the percentage of data that is suboptimal, using the released MH dataset from Robomimic Square. Left: the gain over DP increases as more suboptimal data enters the mixture. The gain is no larger at 100 % suboptimal demonstrations, where there are no optimal suffixes to stitch into. Right: the distribution of lengths of the demonstrations in each dataset. Grey shows the recorded lengths, before stitching. Blue shows each demonstration's shortest route to success after stitching, when one accepted bridge may replace part of it: with fast demonstrations to bridge into, the slow ones shorten to the fast route.

Success rate varying total # of demos ↑ Higher Is Bettermean ± standard error over five seeds

Figure 3.2: Holding the same mixture of optimal vs suboptimal data, we vary the total number of successful demonstrations used to train the policy. The gain from stitching does not disappear even as the number of total demonstrations changes.

4. Experiments on dynamic and more diverse tasks.

Reviewer voED

NEEDLEwork improves success even on non-quasistatic tasks. We test our method on a third-party environment, RoboDyna, where a robot places a cup onto a moving belt.

Success rate ↑ Higher Is Bettermean ± standard error over five seeds

Mean steps per successful trial ↓ Lower Is Bettermean ± standard error over five seeds

Figure 4.1: NEEDLEwork succeeds more often and uses fewer actions among successful trials.
Example: NEEDLEwork finishes sooner than DP.
Example: NEEDLEwork succeeds; DP does not.
Back to contents ↑

5. Composition and total percentage of stitched dataset.

Reviewer ak1N

Different tasks contain different compositions of stitched behavior. Overall, stitches add a small, but important percentage of the training data. Across the three Robomimic tasks, bridges consist of 19.6, 8.3, and 6.5% of the training windows on Can, Square, and Transport, respectively. We provide a more thorough breakdown of behavior below. Click on each bar to see examples. The main Robomimic policies are trained on 254, 245, and 441 bridges on Can, Square, and Transport, respectively.

Contact during training bridges

Figure 5.1: Bridges cover free-space motion and contact: 75 % of Transport bridges hold the object throughout, and 30 % of Can and 17 % of Square bridges involve contact (touching, holding, picking up, or releasing). Click on each segment to see examples.

▸ Click a bar segment above to watch an example bridge of that kind.

Isolated sources. Yes, this happens: 12 / 18 / 3 of the 40 successful demonstrations on Can / Square / Transport are never a bridge source. These demonstrations are already short (median 109 vs 420, 134 vs 540 and 413 vs 940 steps; Figure 5.2), so they offer little to shorten.

Lengths of the successful demonstrations

Figure 5.2: Demonstrations that never become a bridge source are consistently among the shortest of their task. One dot per successful demonstration (40 per task).

Steps remaining before and after each bridge

Figure 5.3: Every bridge lands on a state with fewer steps remaining than its source; the distance below the diagonal is the number of steps saved.

6. Are the three categories needed?

Reviewer 4XGn

The three stages serve different purposes, and combining them gives the highest mean success in the reported subset comparison. We ablate the effects of including different subsets of the three stages, and find that while most subsets improve above DP, including all 3 still helps the most.

The three categories are the three ways a useful target can relate to a source, given that the target must lie on a successful trajectory: on the source's own successful trajectory (S1), on another successful trajectory from a successful source (S2), or on a successful trajectory from a failed source (S3). All three use the same retrieval, IDM, verifier, and selection rule. In earlier development, to test whether the categories themselves are needed, we replaced them with a single ranking of all accepted bridges by verifier score. This recovered part of the gain but fell short of the full method. The key role of the categories is therefore how they allocate the bridge budget: each category supplies a different kind of connection, and the ablation above shows that no single category, and no subset we tested, matches using all three.

Which kinds of connection help? → Higher Is Bettermean ± standard error over five seeds

Figure 6.1: S1 helps skip inefficient/redundant actions within a trajectory. S2 encourages policies to find a new, more efficient way to complete the task from a given starting state. S3 adds additional data coverage from failure demonstrations.

7. How are target states retrieved?

Reviewer ak1N

Retrieving targets for one Square source state

  1. Input: recorded observations, horizon H, candidate budgets
  2. Calibrate distance cutoff r from recorded H-step transitions.
  3. For each source s:
  4. collect eligible targets across trajectories;
  5. retain targets within cosine distance r of s;
  6. sample up to k targets round-robin across distance bins;
  7. propose bridges with the IDM, and keep those the verifier accepts.
  8. Output: verified bridges for policy training.

φ: frozen visual features the policy is trained on. Cosine distance d(a, b) = 1 − cos(φ(a), φ(b)).

Figure 7.1: Retrieval for one source state in the Square dataset. Trajectories are drawn in a 2D projection of the frozen visual features φ; in the rings, distance from the centre is the cosine distance to the source. Click a line of the algorithm to jump to it.

Indeed, exhaustive sampling would provide the broadest coverage of candidate pairs. However, generating and verifying bridges for all O(N2) source–target pairs is expensive, so we design our sampling procedure to distribute a limited budget across sources and visual distances.

First, we exclude target states that cannot provide the required step saving (only S1 and S2 categories, rather than failure-to-success recovery). Retrieval then uses cosine similarity in the frozen visual features the policy is trained on. Since the inverse dynamics model has a fixed horizon of H, we calibrate a visual-distance cutoff from recorded observations H steps apart. For S2 and S3, eligible target states are pooled across trajectories. We do not select a target trajectory first; target states are selected directly from all eligible states. We sample round-robin across visual-distance bins for each source, then allocate the overall candidate budget round-robin across source states with available targets.

8. What does offline rewriting cost?

Reviewer 4XGn

GPU wall-clock for full pipeline. Our method requires just a one-time step to rewrite the offline dataset, and does not impact the actual duration of policy training, nor inference time. All times measured on RTX PRO 6000 Blackwell GPUs. The prepared corpus can be reused across downstream seeds and configurations, so stitching is a one-time cost.

How the cost scales. Rewriting time is dominated by scoring candidates with the IDM and verifier, and it grows linearly with the number of candidates (a least-squares fit gives R² = 0.998 on Can and 0.999 on Square, over 5k–200k candidates). Each source receives a fixed candidate budget, so the cost grows linearly with dataset size rather than with the number of source–target pairs. The GPU time per candidate (eight IDM samples plus verification) is similar across tasks: 7–15 ms on Robomimic and 7–9 ms on the real-robot data.

Square (simulation)

Sweater Folding (real robot)

Figure 8.1: We show a case study on Square (sim) and Sweater Folding (real) on the cost of running our method: training the IDM and Verifier, which can be trained in parallel on the same GPU, and rewriting are one-time costs, shown next to one policy training run.