PACE
NeurIPS 2026 · vision-language-action

PACE

Progress Actively-internalized Conditioning Execution for Long-horizon Manipulation

Fei Ni, Zhuo Chen, Xianze Yao, Zerui Chen, Changrui Chen, Yifu Yuan, Shan Luo, Jianye Hao, Jiankang Deng

Imperial College London · King's College London

Paper · coming soon Code Models BibTeX
pcur

Reads its progress. From the image and the instruction alone, the policy states how far the task has come.

Δpcur

Commits to the next step. Before each action chunk it states how much that chunk should advance the task, and acts on it.

Δpaudit

Checks itself. A frozen encoder measures what the chunk delivered; unmet commitments add up to a failure alarm.

Released checkpoints, raw readings with no smoothing. Before each chunk the policy commits to how far it will advance: blue, dashed until the chunk has run, then part of the committed line. Orange is the realized progress (anchored progress, measured with the frozen encoder). At every check the gap between the two (red: unmet, green: exceeded) drops into the meter of unmet commitments, which is the failure alarm. The window above the curve zooms in on the latest chunks.

PACE overview: foresight-anchored progress pretraining, perceive history and commit to the future, and gains in simulation and on real robots
Method

Progress as an interface inside the policy

PACE turns task progress into something the policy predicts, commits to and conditions on, all within one forward pass of a Qwen3-VL-4B backbone with a flow-matching action head.

  1. Foresight anchor. Learnable goal queries predict an embedding of the finished task in a frozen DINOv2 space. Progress is the projection of the current frame onto the start-to-anchor direction, so no goal image is needed at inference.
  2. Perceive and commit. Two more queries read out the current progress pcur and the forward commitment Δpcur; the increment the last chunk actually achieved, Δpprev, is fed back as a token.
  3. Condition. The anchor, pcur and Δpcur are prepended to the action decoder, so progress shapes the action chunk.
  4. Verify. Realized minus committed increment gives Δpaudit, which drives an online failure alarm. It reports and never changes the actions.
PACE framework diagram
Results

One checkpoint per benchmark

LIBERO (Franka)
Spatial99.0
Object99.8
Goal99.4
Long96.2
Average98.6

Success rate (%) of the released LIBERO checkpoint with the evaluation scripts in the repository: 10 tasks × 50 trials per suite, seed 7.

Progress, live

Eight tasks, one policy reading its own progress

Eight rollouts side by side with their committed and realized progress
Blue: committed progress, orange: realized progress; on successful rollouts the two stay close. The policy's current-progress reading tracks hindsight progress closely: median per-episode Spearman ρ 0.96 on LIBERO.
Failure alarm

It knows when it is failing

Every chunk is a promise the frozen encoder can check. The alarm adds up the progress the policy promised but did not deliver, and watches for progress that stops rising. It needs no extra model and no failure labels.

92.9%of failed LIBERO episodes flagged
7.2%of successful episodes flagged
½of the episode typically still left at the alarm
Citation 
@inproceedings{ni2026pace,
  title     = {{PACE}: Progress Actively-internalized Conditioning Execution for Long-horizon Manipulation},
  author    = {Ni, Fei and Chen, Zhuo and Yao, Xianze and Chen, Zerui and Chen, Changrui and Yuan, Yifu and Luo, Shan and Hao, Jianye and Deng, Jiankang},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}