cs.ROOct 15, 2026
SaveEmbodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations
Organizations: UC Berkeley
Abstract
Flow-based policies such as are trained on a fixed set of embodiments, and transferring them to new ones is an open problem. To adapt a frozen to a new embodiment without additional demonstrations or parameter updates, we train a lightweight embodiment classifier to distinguish successful action chunks on the source body from those on the target body, using a few calibration tasks. On transfer tasks that saw on its source body but the classifier was never trained on, we nudge each denoising step of with the classifier's gradient, which estimates the direction from the distribution of successful actions on the source body toward that on the target body. On the LIBERO simulation benchmark with four modified controllers, embodiment-classifier guidance raises 's average success on the target bodies from 72% to 80% across 12 suite-controller pairs, with a 31% relative reduction in failures (11 of 12 gain; 5 significant after Holm correction). In simulation, guidance is data-efficient, needing few target-body rollouts per task, which makes it promising for sim-to-real transfer, where rollouts are costly. Guidance is mixed on weak wrists and harmful on a locked wrist, when the classifier lacks data or favors infeasible motions. Collecting more target-body data with guided can fill such gaps: on one weak-wrist arm, guided solved a task on which unguided had zero successes in the original data, and retraining the classifier on the guided successes raised success from 52.5% to 70.7%.
Figures & tables
Figure 1: Embodiment-classifier guidance steers a frozen toward plans that succeed on a target body. , trained on a Panda, runs on a KUKA arm with 25% wrist strength. Middle (schematic): the classifier nudges every denoising step. Right: a selected scene where guidance succeeds and unguided fails.
Figure 2: Transfer-task gain over unguided (percentage points; 95% paired-bootstrap CIs). A: LIBERO-Object by body family (squares: registered sweep, where Ctrl(0.75, 0.45 s) is below the registered floor and reported only; diamond: 3-seed mean). B: other suites (hollow: second runs, i.e., full repeats with new target-body data). C: KUKA arm (25% wrist), classifier retrained with extra successes from unguided or guided with the same number of extra target-body queries (bars: 3-seed means; dots: seeds).
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Suite (limit) | Task | Instruction |
|---|---|---|
| LIBERO-Object | 1 (calibration) | pick up the cream cheese and place it in the basket |
| 560 steps | 2 (calibration) | pick up the salad dressing and place it in the basket |
| (native 280) | 4 (calibration) | pick up the ketchup and place it in the basket |
| 5 (calibration) | pick up the tomato sauce and place it in the basket | |
| 0 (transfer) | pick up the alphabet soup and place it in the basket | |
| 3 (transfer) | pick up the bbq sauce and place it in the basket |
Table 1: Tasks and their exact instructions. Calibration tasks supply target-body data for the classifier and the choice of ; all reported results are on transfer tasks, which have no target-body data. was trained on all tasks with the source body. The splits were fixed before any data existed, and code guards block transfer task ids from collection and selection.
| Body | Exact change from the source body |
|---|---|
| source body, Ctrl(1, 0 s) | stock Panda arm, LIBERO’s OSC controller at standard stiffness, no delay |
| Controller changes (Panda; camera images unchanged). Ctrl( , s): stiffness scaled by , commands held back by seconds | |
| Ctrl(1, 0.2 s) | standard stiffness; commands delayed 4 control steps |
| Ctrl(0.5, 0.2 s) | stiffness halved; commands delayed 4 steps |
| Ctrl(0.5, 0.3 s) | stiffness halved; commands delayed 6 steps |
| Ctrl(0.33, 0.3 s) | stiffness one third; commands delayed 6 steps |
Table 2: The target bodies. Stiffness is the gain of the end-effector controller; a delay holds each command back for a fixed number of 20 Hz control steps (2, 3, 4, 6 and 9 steps are 0.1, 0.15, 0.2, 0.3 and 0.45 s).
Figure 3: Toy illustration with made-up 2-D data (not robot data). Each point is a 2-D “action”. The source body’s actions form two clusters (dashed circles); the target body succeeds only in the shaded region. Lines show 60 samples moving from noise ( ) to actions ( ) under the exact flow of the source-body data. Left: the frozen flow puts 48% of samples in the shaded region. Right: adding a capped classifier gradient (here a fixed linear logit that prefers larger ) puts 63% there. Guidance mostly moves samples toward the source-body cluster that also works on the target body.
| LIBERO-Object, 560 steps, transfer tasks 0, 3, 6, 7 | |||
|---|---|---|---|
| Body | Ung. % (n) | Guid. % (n) | Diff, points [95% CI] |
| Ctrl(0.5, 0.3 s) | 73 (146) | 93 (186) | +20 [+13.5, +27] |
| classifier seed 1 | 73 (146) | 93.5 (187) | +20.5 [+13.5, +27.5] |
| classifier seed 2 ( ) | 73 (146) | 91 (182) | +18 [+11, +25] |
| eval seed 1 | 72 (144) | 96 (192) | +24 [+17.5, +30.5] |
| Ctrl(0.33, 0.3 s) | 88.5 (177) | 97.5 (195) | +9 [+4, +14] |
Table 3: Transfer-task success, unguided (Ung.) vs. guided (Guid.) , in % with successes out of 200 in parentheses. One evaluation seed and classifier per row unless stated. KUKA arm: classifier seeds give 57% (114/200), 52% (104/200) and 48.5% (97/200); the guided-collection row is the 3-seed mean (App. C.3 ). 2nd: the gain of a second run with fresh target-body data, a new classifier and a new ; pool: the first run’s gain averaged over two evaluation seeds (all in points). † Registered after the first run was inconclusive; these second runs go from 43.5% (87/200) to 57% (114/200) and from 61% (122/200) to 73.5% (147/200). L-10: LIBERO-10. The sweep bodies are in Table 6 .
Figure 4: Gain vs. unguided success for every first run in Table 3 and the six evaluated sweep bodies (Table 6 ), with 95% paired-bootstrap CIs. The dashed line is the largest possible gain. Gains appear across a wide range of starting success; near the ceiling there is no room, and the largest drops sit at low or middle success.
| Body | Task 0 | Task 3 | Task 6 | Task 7 | Task 1 | Total % (n/200) | Diff, points [95% CI] |
|---|---|---|---|---|---|---|---|
| Ctrl(0.5, 0.3 s) | 78 / 94 | 52 / 84 | 88 / 100 | 74 / 94 | 82 / 90 | 73 / 93 (146 / 186) | +20 [+13.5, +27] |
| Ctrl(0.5, 0.3 s) (eval seed 1) | 76 / 100 | 48 / 94 | 96 / 100 | 68 / 90 | 76 / 84 | 72 / 96 (144 / 192) | +24 [+17.5, +30.5] |
| Ctrl(0.33, 0.3 s) | 94 / 100 | 72 / 92 | 92 / 100 | 96 / 98 | 86 / 100 | 88.5 / 97.5 (177 / 195) | +9 [+4, +14] |
| Ctrl(1, 0.2 s) | 82 / 90 | 56 / 76 | 98 / 100 | 72 / 74 | 76 / 82 | 77 / 85 (154 / 170) | +8 [+1, +15] |
| Ctrl(1, 0.2 s) (eval seed 1) | 80 / 92 | 40 / 76 | 94 / 100 | 76 / 84 | 72 / 72 | 72.5 / 88 (145 / 176) | +15.5 [+8, +23] |
| UR5e (speed cap) | 80 / 80 | 94 / 78 | 72 / 66 | 90 / 86 | 100 / 92 | 84 / 77.5 (168 / 155) | 6.5 [ 13, 0] |
Table 4: Per-task results, LIBERO-Object (560 steps, starting states 0–49). Cells: unguided / guided success in % of 50 episodes; bold: guided at least 10 points above unguided. Grey: calibration task 1 (not a transfer task). Totals are over transfer tasks 0, 3, 6 and 7 (200 episodes), in % with counts in parentheses, and the paired difference in points with its 95% paired-bootstrap CI. ‡ Outside the pilot band (App. D ).
| Suite | Body | Task 1 | Task 3 | Task 5 | Task 7 | Task 0 | Total % (n/200) | Diff, points [95% CI] |
|---|---|---|---|---|---|---|---|---|
| LIBERO-Goal | Ctrl(0.5, 0.3 s) | 84 / 92 | 18 / 38 | 42 / 40 | 98 / 98 | 96 / 82 | 60.5 / 67 (121 / 134) | +6.5 [ 0.5, +14] |
| LIBERO-Goal | Kinova arm (10% wrist) | 66 / 66 | 22 / 24 | 6 / 10 | 88 / 88 | 0 / 0 | 45.5 / 47 (91 / 94) | +1.5 [ 6, +8.5] |
| LIBERO-Spatial | Ctrl(0.5, 0.3 s) | 60 / 92 | 20 / 58 | 46 / 40 | 94 / 98 | 54 / 38 | 55 / 72 (110 / 144) | +17 [+9, +25.5] |
| LIBERO-10 | Ctrl(0.5, 0.3 s) | 76 / 80 | 92 / 86 | 86 / 84 | 72 / 54 | 48 / 70 | 81.5 / 76 (163 / 152) | 5.5 [ 13, +2] |
| LIBERO-Goal | KUKA arm (25% wrist) | 76 / 66 | 28 / 30 | 2 / 12 | 74 / 100 | 0 / 6 | 45 / 52 (90 / 104) | +7 [0, +14] |
| LIBERO-Goal | Ctrl(1, 0.2 s) | 76 / 88 | 24 / 38 | 32 / 34 | 100 / 98 | 96 / 96 | 58 / 64.5 (116 / 129) | +6.5 [ 0.5, +13.5] |
Table 5: Per-task results, other suites (2 the native horizon; transfer tasks 1, 3, 5 and 7). Cells as in Table 4 ; grey: calibration task 0. † One run, recomputed after the other rows; unguided is already at 93%, so by the registered rule this body has no room to gain.
| Suite | Target (no data) | Classifier from | Ung. | Reused | Own | Reused ung. | Reused own |
|---|---|---|---|---|---|---|---|
| Object | Ctrl(0.5, 0.3 s) | Ctrl(0.5, 0.2 s) | 73 (146) | 96 (192) | 93 (186) | +23 [+16.5, +29.5] | +3 [ 1.5, +7.5] |
| Object | Ctrl(0.5, 0.2 s) | Ctrl(0.5, 0.3 s) | 94.5 (189) | 97 (194) | 97.5 (195) | +2.5 [ 1.5, +6.5] | 0.5 [ 4, +3] |
| Goal | Ctrl(0.33, 0.3 s) | Ctrl(0.5, 0.3 s) | 72.3 (144.5) a | 82 (164) | 81.8 (163.5) b | +9.75 [+3.75, +15.5] | +0.25 [ 5.75, +6] |
| Object | Kinova (10% wrist) | KUKA (25% wrist) | 70 (140) | 58.5 (117) | 85 (170) | 11.5 [ 19.5, 4] | 26.5 [ 34.5, 18.5] |
| Object | Kinova (10% wrist) | KUKA, guided coll. | 70 (140) | 65.5 (131) | 85 (170) | 4.5 [ 13.5, +4.5] | 19.5 [ 27.5, 11.5] |
| Object | Panda (10% wrist) | KUKA (25% wrist) | 47 (94) | 24.5 (49) | 37.5 (75) | 22.5 [ 31, 14] | 13 [ 21, 5] |
Table 7: Reusing a classifier on a body that contributed no data. Success in % with successes out of 200 in parentheses; differences in points. Top: sibling controller bodies (the second row is at the ceiling). Bottom: across arms. Own: the target’s own classifier; guided coll.: the KUKA arm’s guided-collection classifier. a Mean of unguided runs 69.5% (139/200) and 75% (150/200). b Mean of its two classifiers, 81.5% (163/200) and 82% (164/200).
More in this topic
Cross-Embodiment Transfer
Cross-embodiment transfer moves skills and learned representations between robots with different bodies, sensors, and action spaces.
See all 43 papers