Latent Action Policies from Suboptimal Data
Organizations: Korea University · Microsoft Research Asia
Abstract
Latent Action Models (LAMs) learn control representations from action-free video, but existing pipelines typically train latent policies with behavioral cloning, without an explicit policy improvement objective. Although offline RL offers a natural approach to policy improvement, standard methods do not transfer directly to learned latent action spaces. We find that the structure of learned latent actions makes naive offline RL unreliable: directly maximizing value over latent actions or fitting a unimodal policy can produce latent actions that cannot be realized by physical execution, while the latent representation can merge physical controls that receive different rewards. To understand when latent-space optimization remains valid, we derive conditions under which policy improvement transfers to decoded execution and under which latent and physical optima align. Building on these insights, we propose LAPIS for goal-conditioned latent policy improvement, grounding value-based policy improvement in observed state-latent transitions and using a conditional flow policy to capture their multimodal structure. LAPIS outperforms baselines on most tasks, with the clearest gains in visual manipulation, and remains effective even with only 1% action labels, where competing methods largely collapse.
Figures & tables
| labels (oracle) | labels (w/ LAM) | |||||||||||||||
| Task | GCBC | F-GCBC | GCIVL | GCIQL | QRL | HIQL | CRL | LAPIS | F-GCBC | GCIVL | GCIQL | QRL | HIQL | CRL | LAPIS | LAPIS † |
| antmaze-medium | 12 | 8 | 23 | 12 | 0 | 94 | 95 | 48 | 8 | 26 | 7 | 34 | 89 | 11 | 49 | 89 |
| antmaze-large | 4 | 1 | 6 | 3 | 1 | 66 | 85 | 14 | 1 | 6 | 2 | 9 | 49 | 3 | 10 | 58 |
| antmaze-giant | 0 | 0 | 1 | 0 | 1 | 8 | 47 | 0 | 0 | 1 | 0 | 0 | 3 | 0 | 0 | 4 |
| antmaze-teleport | 5 | 7 | 8 | 7 | 10 | 42 | 51 | 23 | 6 | 6 | 4 | 14 | 34 | 6 | 21 | 39 |
| cube-single | 4 | 11 | 58 | 30 | 28 | 86 | 18 | 45 | 4 | 45 | 5 | 5 | 56 | 4 | 62 | 33 |
| labels (oracle) | labels (w/ LAM) | |||||
|---|---|---|---|---|---|---|
| Task | GCIQL | CRL | LAPIS | GCIQL | CRL | LAPIS |
| partial | 21 | 28 | 60 | 19 | 26 | 44 |
| mixed | 22 | 31 | 53 | 21 | 25 | 41 |
| labels | labels | |||||||||
| Task | SS-ORL | VPT-RL | GCIVL | HIQL | LAPIS | SS-ORL | VPT-RL | GCIVL | HIQL | LAPIS |
| antmaze-medium | 23 | 11 | 2 | 42 | 49 | 0 | 0 | 0 | 0 | 29 |
| antmaze-large | 6 | 4 | 0 | 7 | 10 | 0 | 0 | 0 | 0 | 7 |
| antmaze-giant | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| antmaze-teleport | 13 | 8 | 3 | 12 | 21 | 4 | 3 | 0 | 2 | 11 |
| cube-single | 55 | 48 | 0 | 24 | 62 | 0 | 0 | 0 | 2 | 50 |
| Improvement | Actor | antmaze-medium | cube-single | scene |
|---|---|---|---|---|
| CRL (max- ) | det | 11 | 4 | 6 |
| CRL (max- ) | flow | 5 | 5 | 7 |
| LAPIS (gate) | det | 43 | 36 | 16 |
| LAPIS (gate) | flow | 49 | 62 | 29 |
| Env | Decoder | labels | labels | labels |
|---|---|---|---|---|
| cube | 54 | 62 | 50 | |
| 74 | 55 | 46 | ||
| scene | 25 | 29 | 24 | |
| 38 | 20 | 16 | ||
| antmaze | 48 | 49 | 44 | |
| 52 | 27 | 38 |
| antmaze | cube | scene | |
|---|---|---|---|
| Spearman | |||
| Gate agreement |
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
| Search | Population iterations | Interpolation | State swap |
|---|---|---|---|
| Default | |||
| Enlarged |
| Setting | AntMaze-medium | Cube-single | Scene | ||
|---|---|---|---|---|---|
| One-step representation | 1 | 1 | |||
| LAPIS | 3 | 1 | |||
| Full-chunk execution | 3 | 3 |
| Environment | Subset | |||
|---|---|---|---|---|
| v-cube-single-play | All | 100 | 3.26 | 1.51 |
| Moving | 42 | 7.69 | 3.59 | |
| v-scene-play | All | 100 | 1.22 | 0.80 |
| Moving | 19 | 6.42 | 4.20 |
| Setting | Visual manipulation | Visual antmaze |
|---|---|---|
| Adam learning rate | ||
| Batch size | ||
| Gradient steps | k | k |
| Visual encoder | small IMPALA | small IMPALA |
| Policy MLP trunk | ||
| Layer normalization | yes | yes |
| Setting | Visual manipulation | Visual antmaze |
|---|---|---|
| BC coefficient |
| Task | Standard candidates | Grid-restricted candidates |
|---|---|---|
| v-cube-single-play | ||
| v-scene-play | ||
| v-antmaze-medium-navigate |
| Method | AntMaze-medium | Cube-single | Scene | |
|---|---|---|---|---|
| GCIQL | 1 | |||
| GCIQL | 3 | |||
| CRL | 1 | |||
| CRL | 3 | |||
| LAPIS | 3 |
| Task | |||
| v-antmaze-medium-navigate | |||
| v-antmaze-large-navigate | |||
| v-antmaze-giant-navigate | |||
| v-antmaze-teleport-navigate | |||
| v-cube-single-play | |||
| v-cube-double-play |
| Task | Action space | |||
|---|---|---|---|---|
| v-antmaze-medium-navigate | Real | |||
| Latent | ||||
| v-cube-single-play | Real | |||
| Latent | ||||
| v-scene-play | Real | |||
| Latent |
| Task | Overall | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| v-antmaze-medium-navigate | |||||||||
| v-antmaze-large-navigate | |||||||||
| v-antmaze-giant-navigate | |||||||||
| v-antmaze-teleport-navigate | |||||||||
| v-cube-single-play | – | – | – | ||||||
| v-cube-double-play | – | – | – |
| Task | Overall | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| D4RL Kitchen |
| Environment | ||
|---|---|---|
| v-antmaze-medium-navigate | ||
| v-cube-single-play | ||
| v-scene-play |
| Task | Dataset latents | LAPIS ( ) | AWR-flow | max- |
|---|---|---|---|---|
| cube-single | ||||
| scene |
| cube-single | scene | |
|---|---|---|
| Failure rate (%), with / without projection | / | / |
| Success difference, without with (points) |
| Task | F-GCBC | GCIVL | GCIQL | QRL | HIQL | CRL | LAPIS |
|---|---|---|---|---|---|---|---|
| antmaze-medium | 3 | 3 | 4 | 0 | 8 | 8 | 2 |
| antmaze-large | 1 | 1 | 0 | 0 | 2 | 0 | 0 |
| antmaze-giant | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| antmaze-teleport | 2 | 1 | 1 | 1 | 3 | 1 | 2 |
| cube-single | 2 | 5 | 7 | 8 | 5 | 1 | 4 |
| cube-double | 0 | 0 | 0 | 0 | 1 | 0 | 0 |
Latent Action Learning
Latent action learning discovers compact representations of behavior from observations or trajectories to support control, planning, and transfer.
See all 125 papers