SAMBAR

Overview

An overview of SAMBAR: a replay-free continual learning algorithm that prevents catastrophic forgetting during VLA finetuning without requiring access to the demonstrations of any previously learned task. A policy learns tasks $T_1, T_2, T_3$ sequentially using only the current task's demonstrations. Orange, blue, and green indicate parameter importance for $T_1$, $T_2$, and $T_3$, with darker shades indicating greater importance. Standard supervised finetuning (SFT) overwrites parameters important to earlier tasks, causing forgetting. SAMBAR limits changes to those parameters using method of multipliers along with importance weighting while allowing other parameters to adapt to new tasks.

Hardware Experiments

SFT and SAMBAR performance on an xArm7 manipulator. Videos play at 10× speed.

Both policies are finetuned on cake-on-cake-tray and then on remove-lid-and-place-bird, in sequence. SAMBAR remembers the pretraining task wheras SFT forgets. Both have similar performance on the both finetuning tasks.

eraser-in-bowl

SFT

SAMBAR (ours)

cake-on-cake-tray

SFT

SAMBAR (ours)

remove-lid-and-place-bird

SFT

SAMBAR (ours)

Both policies are finetuned on cake-on-cake-tray only, starting from the four pretraining tasks. SFT forgets the pretraining task: two-birds-in-binwhereas SAMBAR remembers all the pretraining tasks.

two-birds-in-bin

SFT

SAMBAR (ours)

cake-on-cake-tray

SFT

SAMBAR (ours)

Both policies are finetuned on remove-lid-and-place-bird only, starting from the four pretraining tasks. SFT forgets the pretraining task: eraser-in-mugwhereas SAMBAR remembers all the pretraining tasks.

eraser-in-mug

SFT

SAMBAR (ours)

remove-lid-and-place-bird

SFT

SAMBAR (ours)

Sequential continual learning on hardware

SAMBAR is the only method that remembers all four pretraining tasks while acquiring both new tasks at 93.3% Weighted Avg SR against 86.7% for the best baseline.

Show the exact numbers

Method	Pretraining tasks ↑				Pretrain SR ↑	Finetune tasks ↑		Weighted Avg SR ↑
Method	bird-in-bin	two-birds-in-bin	eraser-in-mug	eraser-in-bowl	Pretrain SR ↑	cake-on-cake-tray	remove-lid-and-place-bird	Weighted Avg SR ↑
SFT	5/5	5/5	5/5	3/5	90.0	4/5	3/5	83.3
RETAIN ($\alpha=0.5$)	5/5	5/5	4/5	4/5	90.0	5/5	0/5	76.7
EWC	5/5	5/5	5/5	5/5	100.0	4/5	2/5	86.7
SAMBAR (ours)	5/5	5/5	5/5	5/5	100.0	4/5	4/5	93.3

Sequential continual learning on hardware. Individual tasks are reported as the number of successful rollouts out of 5. Pretrain SR is the percentage of the 20 pretraining rollouts; Weighted Avg SR is the percentage of all 30 rollouts.

Finetuning on a single task

SAMBAR is the only method that have high performance on every task at 100% Weighted Avg SR against 88.0% for the best baseline.

Show the exact numbers

Method	Pretraining tasks ↑				Pretrain SR ↑	Finetune SR ↑	Weighted Avg SR ↑
Method	bird-in-bin	two-birds-in-bin	eraser-in-mug	eraser-in-bowl	Pretrain SR ↑	Finetune SR ↑	Weighted Avg SR ↑
Finetune task: cake-on-cake-tray
SFT	4/5	1/5	5/5	2/5	60.0	5/5	68.0
RETAIN ($\alpha=0.5$)	5/5	4/5	3/5	5/5	85.0	5/5	88.0
EWC	5/5	4/5	5/5	2/5	80.0	4/5	80.0
SAMBAR (ours)	5/5	5/5	5/5	5/5	100.0	5/5	100.0
Finetune task: remove-lid-and-place-bird
SFT	5/5	4/5	4/5	3/5	80.0	5/5	84.0
RETAIN ($\alpha=0.5$)	5/5	5/5	3/5	5/5	90.0	0/5	72.0
EWC	5/5	5/5	3/5	5/5	90.0	0/5	72.0
SAMBAR (ours)	5/5	5/5	5/5	5/5	100.0	4/5	96.0

Learning one new task on hardware. We pretrain $\pi_{0.5}$ on four tasks and then finetune it on each of the two finetuning tasks separately. Every task is evaluated with 5 rollouts. Pretrain SR is the percentage of the 20 pretraining rollouts that succeed; Weighted Avg SR is the percentage of all 25 rollouts.

Simulation Experiments

The three tasks we learn in sequence

pot-on-stove $T_1$

"turn on the stove and put the moka pot on it"

bowl-in-drawer $T_2$

"put the black bowl in the bottom drawer of the cabinet and close it"

bottle-on-rack $T_3$

"put the wine bottle on the rack"

Starting from a policy pretrained on 32 LIBERO tasks, we finetune on pot-on-stove, then bowl-in-drawer, then bottle-on-rack, in the order shown. The first two are long-horizon tasks; the third is a goal task.

After learning all three tasks

SAMBAR performs every one of them, while SFT has has completely forgotten first task.

SAMBAR (Ours)

pot-on-stove ($T_1$)

bowl-in-drawer ($T_2$)

bottle-on-rack ($T_3$)

SFT

pot-on-stove ($T_1$) fully forgotten

bowl-in-drawer ($T_2$) forgotten

bottle-on-rack ($T_3$)

Rollouts of the final policies, played at 2× speed. SFT succeeds on pot-on-stove 0.0% of the time after learning the two later tasks, while SAMBAR still performs it 84.0% of the time.

Sequential Continual Learning

Every baseline falls to 0.0% on pot-on-stove after finetuning on two more tasks; SAMBAR holds 84.0% and leads Weighted Avg SR at 87.4% against 56.7% for the best baseline.

Show the exact numbers

Method	Pretrain SR ↑	Success on each newly learned task ↑			Weighted Avg SR ↑
Method	Pretrain SR ↑	pot-on-stove	bowl-in-drawer	bottle-on-rack	Weighted Avg SR ↑
Co-Training	95.4 ±1.0	100.0 ±0.0	98.0 ±5.6	88.0 ±13.6	95.4
SFT	12.9 ±1.9	0.0 ±0.0	60.0 ±21.5	84.0 ±11.1	15.9
RETAIN	17.4 ±0.7	0.0 ±0.0	68.0 ±16.2	96.0 ±6.8	20.6
LoRA	45.4 ±1.3	0.0 ±0.0	18.0 ±13.6	92.0 ±10.4	44.6
EWC	0.0 ±0.0	0.0 ±0.0	0.0 ±0.0	100.0 ±0.0	2.9
L2	58.3 ±2.6	0.0 ±0.0	42.0 ±34.5	80.0 ±24.8	56.7
Simple Recipe Works	44.4 ±3.2	0.0 ±0.0	36.0 ±14.2	90.0 ±8.8	44.2
SAMBAR w/o importance	89.8 ±0.8	0.0 ±0.0	0.0 ±0.0	4.0 ±6.8	82.2
SAMBAR (ours)	88.2 ±2.3	84.0 ±6.8	70.0 ±15.2	82.0 ±16.2	87.4

Sequential continual learning on LIBERO. Success rates in percent, averaged over 5 seeds with 10 episodes/seed; error bars are 95% confidence intervals, clipped at 100%. Higher is better (↑) everywhere. Weighted Avg SR is the task-count-weighted mean over all 35 tasks.

Single-Task Finetuning

Narrowly pretrained (32 tasks)

SAMBAR retains 89.7% of the pretraining tasks while still learning the new one at 86.0%; the best baseline retains only 84.0%.

Show the exact numbers

Method	Pretrain SR ↑	Finetune SR ↑	Weighted Avg SR ↑
Finetune task: pot-on-stove
Pretrained (no finetuning)	93.9 ±1.4	6.0 ±16.7	–
Co-Training	93.8 ±1.3	94.0 ±6.8	93.8
SFT	56.1 ±2.6	96.0 ±6.8	57.3
RETAIN	66.5 ±2.1	94.0 ±11.1	67.3
LoRA	75.4 ±1.8	100.0 ±0.0	76.2
EWC	47.1 ±1.4	96.0 ±11.1	48.6
L2	71.5 ±2.2	74.0 ±16.7	71.6
Simple Recipe Works	84.0 ±1.3	86.0 ±16.7	84.0
SAMBAR w/o importance	91.8 ±1.4	4.0 ±6.8	89.1
SAMBAR (ours)	89.7 ±1.7	86.0 ±18.8	89.6
Finetune task: bowl-in-drawer
Pretrained (no finetuning)	93.9 ±1.4	0.0 ±0.0	–
Co-Training	93.1 ±1.7	100.0 ±0.0	93.3
SFT	23.1 ±1.4	100.0 ±0.0	25.4
RETAIN	27.4 ±1.8	96.0 ±6.8	29.5
LoRA	74.0 ±3.0	100.0 ±0.0	74.8
EWC	19.6 ±0.9	90.0 ±12.4	21.7
L2	64.9 ±2.3	78.0 ±16.2	65.3
Simple Recipe Works	74.1 ±1.2	92.0 ±10.4	74.5
SAMBAR w/o importance	90.6 ±1.9	38.0 ±30.9	89.0
SAMBAR (ours)	91.8 ±1.2	96.0 ±6.8	91.9

Learning one new task from a narrowly pretrained policy. $\pi_{0.5}$ pretrained on 32 LIBERO tasks (8 from each suite), finetuned on two LIBERO-Long tasks. 5 seeds × 10 episodes/seed; error bars are 95% confidence intervals, clipped at 100%. Weighted Avg SR is over all 33 tasks. Forgetting is severe here: every baseline forgets the pretraining tasks, whereas SAMBAR remembers them.

Broadly pretrained (117 tasks)

Forgetting is mild from a broadly pretrained policy, but SAMBAR still has the highest Weighted Avg SR at 91.9%.

Show the exact numbers

Method	Pretrain SR ↑	Finetune SR ↑	Weighted Avg SR ↑
Finetune task: pot-on-stove
Pretrained (no finetuning)	93.2 ±1.2	4.0 ±11.1	–
SFT	89.0 ±0.9	100.0 ±0.0	89.1
RETAIN	89.7 ±1.2	100.0 ±0.0	89.8
LoRA	88.1 ±0.6	100.0 ±0.0	88.2
EWC	89.5 ±1.4	100.0 ±0.0	89.6
L2	84.8 ±0.7	84.0 ±11.1	84.8
SAMBAR w/o importance	89.7 ±1.4	56.0 ±32.4	89.5
SAMBAR (ours)	91.9 ±0.9	92.0 ±13.6	91.9
Finetune task: items-into-basket
Pretrained (no finetuning)	93.2 ±1.2	0.0 ±0.0	–
SFT	93.2 ±1.5	100.0 ±0.0	93.3
RETAIN	93.3 ±2.1	96.0 ±11.1	93.4
LoRA	93.4 ±0.9	100.0 ±0.0	93.5
EWC	93.5 ±1.2	100.0 ±0.0	93.6
L2	93.8 ±1.2	92.0 ±13.6	93.8
SAMBAR w/o importance	94.3 ±0.5	36.0 ±11.1	93.8
SAMBAR (ours)	94.3 ±1.3	100.0 ±0.0	94.4
Finetune task: mugs-on-plates
Pretrained (no finetuning)	93.2 ±1.2	0.0 ±0.0	–
SFT	80.7 ±2.0	96.0 ±11.1	80.8
RETAIN	85.4 ±1.1	88.0 ±22.2	85.5
LoRA	91.9 ±1.3	100.0 ±0.0	91.9
EWC	78.8 ±0.8	96.0 ±11.1	78.9
L2	86.1 ±0.6	80.0 ±30.4	86.0
SAMBAR w/o importance	93.2 ±1.2	76.0 ±20.8	93.1
SAMBAR (ours)	90.7 ±1.1	92.0 ±13.6	90.8

Learning one new task from a broadly pretrained policy. $\pi_{0.5}$ pretrained on 117 LIBERO tasks. 5 seeds × 5 episodes/seed; error bars are 95% confidence intervals, clipped at 100%. Weighted Avg SR is over all 118 tasks. Forgetting is mild in this regime, and SAMBAR learns each new task without significantly forgetting the 117 pretrained tasks.

BibTeX

@article{shrivastava2026sambar,
  title   = {SAMBAR: Selective Anchoring via Method of Multipliers for Balanced
             Knowledge Acquisition and Retention in Vision-Language-Action Models},
  author  = {Shrivastava, Aayushi and Zhou, Xunlan and Zhao, Hongrui and
             Chen, Ziyu and Mehr, Negar},
  journal = {arXiv preprint arXiv:2609.32108},
  year    = {2026}
}

SAMBAR: Selective Anchoring via Method of Multipliers for Balanced Knowledge Acquisition and Retention in Vision-Language-Action Models

Overview

Hardware Experiments

Sequential continual learning on hardware

Finetuning on a single task

Simulation Experiments

The three tasks we learn in sequence

After learning all three tasks

Sequential Continual Learning

Single-Task Finetuning

Narrowly pretrained (32 tasks)

Broadly pretrained (117 tasks)

BibTeX