Literature
Display
← All papers

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

Aayushi Shrivastava · Xunlan Zhou · Hongrui Zhao · Ziyu Chen · Negar Mehr

Preprint posted · arXiv (preprint) ↗

Robotics8 min text · CondensedOriginal paper ↗
Contents

At a glance

SAMBAR finetunes a vision-language-action policy on new tasks while anchoring parameters important to older tasks. It combines diagonal gradient-based importance estimates with a method-of-multipliers update that accumulates parameter drift.

Sequential LIBERO experiments retain 88.2% pretraining success and 78.7% mean success on three new tasks. In a small hardware study, final success is 20/20 on old-task trials and 4/5 on each of two new tasks. The method reduces forgetting in these settings, with a practical trade-off between retention, acquisition and training budget.

I INTRODUCTION

Finetuning a deployed robot for a new manipulation task can degrade earlier skills. Replay addresses this using old demonstrations; SAMBAR instead retains checkpoints and compact parameter-importance statistics.

The model provider must supply an initial importance estimate for the pretrained tasks, as assumed in this paper. A checkpoint alone does not contain the additional statistic required by the method. The experiments construct pretrained LIBERO policies on which that statistic can be computed.

Complete original Fig. 1: Overview of SAMBAR for replay-free continual VLA learning. A policy learns tasks T 1 , T 2 , T 3 T_{1},T_{2},T_{3} sequentially using only the current task’s demonstrations. Orange , blue , and green indicate parameter importance for T 1 {\color[rgb]{0.8945,0.4219,0.043}T_{1}} , T 2 {\color[rgb]{0.3281,0.5586,0.8359}T_{2}} , and T 3 {\color[rgb]{0.4688,0.5781,0.2344}T_{3}} , respectively,
with darker shades indicating greater importance.
Standard supervised finetuning (SFT) can overwrite parameters important to earlier tasks, causing forgetting.
SAMBAR combines an importance-weighted parameter anchor with the method of multipliers to limit changes to parameters important to prior tasks while allowing new-task adaptation.
Fig. 1: Overview of SAMBAR for replay-free continual VLA learning. A policy learns tasks T1,T2,T3T_{1},T_{2},T_{3} sequentially using only the current task’s demonstrations. Orange, blue, and green indicate parameter importance for T1{\color[rgb]{0.8945,0.4219,0.043}T_{1}}, T2{\color[rgb]{0.3281,0.5586,0.8359}T_{2}}, and T3{\color[rgb]{0.4688,0.5781,0.2344}T_{3}}, respectively, with darker shades indicating greater importance. Standard supervised finetuning (SFT) can overwrite parameters important to earlier tasks, causing forgetting. SAMBAR combines an importance-weighted parameter anchor with the method of multipliers to limit changes to parameters important to prior tasks while allowing new-task adaptation.

II RELATED WORKS

The paper compares replay, parameter isolation, model merging and regularization. Replay stores historical examples; adapters restrict the update space; merging blends models; importance-weighted regularization penalizes changes to selected parameters.

SAMBAR adds a signed dual accumulator to an importance-weighted moving anchor. EWC [5] supplies the related Fisher-diagonal motivation, while the anchor and dual update distinguish the proposed training procedure.

III Problem Formulation

Tasks arrive sequentially as language-conditioned demonstration datasets. Each stage starts from the previous policy and uses only the current task’s demonstrations. The generic behavioral-cloning objective is the mean negative log likelihood of demonstrated actions.

The goal is performance on both the original pretraining tasks and all subsequently learned tasks. Retention, current-task acquisition and a task-count-weighted average are therefore evaluated separately.

Equation 1
LB​C(θ;𝒟i)=−1∑j=1NiHj∑j=1Ni∑t=1Hjlogπθ(at∣ot,Ti).L^{BC}(\theta;\mathcal{D}_{i})=-\frac{1}{\sum_{j=1}^{N_{i}}H_{j}}\sum_{j=1}^{N_{i}}\sum_{t=1}^{H_{j}}\log\pi_{\theta}(a_{t}\mid o_{t},T_{i}). (1)

IV Preliminaries

An augmented Lagrangian adds a linear dual term and a quadratic constraint penalty to the task loss. The dual variable accumulates signed constraint violations and feeds them into later updates; the penalty coefficient remains fixed.

The displayed method-of-multipliers equations use a primal minimization. Algorithm 1 implements an optimizer step at each iteration. The general optimization framework does not itself prove convergence or zero forgetting for this nonconvex neural training procedure [12].

Equations 2–4b
minϕ⁡L⁡(ϕ)s.t.g⁡(ϕ)=0,\min_{\phi}L(\phi)\hskip 14.22636pt\text{s.t.}\hskip 14.22636ptg(\phi)=0, (2)
ℒa:=L⁡(ϕ)+g​(ϕ)′​u+ρ2​‖g⁡(ϕ)‖22,\mathcal{L}_{a}:=L(\phi)+g(\phi)^{\prime}u+\frac{\rho}{2}||g(\phi)||^{2}_{2}, (3)
ϕk\displaystyle\phi^{k} =arg⁡minϕ⁡L⁡(ϕ)+g​(ϕ)′​uk−1+ρ2​‖g⁡(ϕ)‖22,\displaystyle=\arg\min_{\phi}L(\phi)+g(\phi)^{\prime}u^{k-1}+\tfrac{\rho}{2}\|g(\phi)\|^{2}_{2}, (4a)
uk\displaystyle u^{k} =uk−1+ρ​g​(ϕk).\displaystyle=u^{k-1}+\rho\,g(\phi^{k}). (4b)

V SAMBAR

SAMBAR first writes parameter preservation as a constrained training problem, then replaces uniform preservation with a per-parameter importance-weighted anchor. All original equations and the complete algorithm remain below.

V-A Continual Learning as Constrained Optimization Problem

The initial equality constraint fixes every parameter to its previous value. Applied exactly, this would prevent adaptation. Its practical uniform-anchor ablation strongly retains old performance but often fails to acquire new skills.

Finite primal/dual updates provide an optimization mechanism for penalizing drift. Importance weighting is introduced to make that anchoring selective.

Equations 5–6b
minθi⁡LB​C​(θi,𝒟i)s.t.θi=θi−1∗,\min_{\theta_{i}}L^{BC}(\theta_{i};\mathcal{D}_{i})\hskip 14.22636pt\text{s.t.}\hskip 14.22636pt\theta_{i}=\theta^{*}_{i-1}, (5)
θik\displaystyle\theta^{k}_{i} =arg⁡minθi​LB​C​(θi,𝒟i)+(θi−θi−1∗)′​uk−1\displaystyle=\arg\min_{\theta_{i}}L^{BC}(\theta_{i};\mathcal{D}_{i})+(\theta_{i}-\theta^{*}_{i-1})^{\prime}u^{k-1}
+ρ2​‖θi−θi−1∗‖2,\displaystyle\quad+\tfrac{\rho}{2}\|\theta_{i}-\theta^{*}_{i-1}\|^{2}, (6a)
uk\displaystyle u^{k} =uk−1+ρ⁡(θik−θi−1∗).\displaystyle=u^{k-1}+\rho\,(\theta_{i}^{k}-\theta^{*}_{i-1}). (6b)

V-B Importance Weighted Anchor

The anchor blends the current and previous parameters according to new-task and accumulated old-task importance. Old-task-sensitive parameters remain close to their previous values; parameters important primarily to the new task can move further.

New-task importance is estimated once at the start of a stage and then averaged into the stored statistic using task counts. The archive is diagonal, so its size grows linearly with parameter count and remains fixed as more tasks arrive. Training also needs the current/old parameters, anchor, dual values and optimizer state.

Equation 8 describes an average of individual-sample squared gradients, whereas Algorithm 1 squares minibatch gradients and averages over sampled batches. These estimators can differ. The algorithm also includes a numerical stabilizer in the elementwise anchor formula that the main matrix equation omits.

The selectable algorithm preserves all 23 original numbered steps, Require/Ensure headers, optimizer update, dual ascent, stabilized anchor and importance subroutine. Its output success statement describes the intended result.

Equations 7–10
zk\displaystyle z^{k} :=(Ωi+Ω¯i−1)−1​(Ωi​θik+Ω¯i−1​θi−1∗),\displaystyle:=(\Omega_{i}+\bar{\Omega}_{i-1})^{-1}(\Omega_{i}\theta_{i}^{k}+\bar{\Omega}_{i-1}\theta^{*}_{i-1}), (7)
Ωi=diag⁡(1NB​∑j=1NB∇θiLB​C​(θi,bj)​(∇θiLB​C​(θi,bj))′),\Omega_{i}=\mathrm{diag}\Big(\frac{1}{N_{B}}\sum_{j=1}^{N_{B}}\nabla_{\theta_{i}}L^{BC}(\theta_{i};b_{j})\,\big(\nabla_{\theta_{i}}L^{BC}(\theta_{i};b_{j})\big)^{\prime}\Big), (8)
θik\displaystyle\theta^{k}_{i} =arg⁡minθi​LB​C​(θi,𝒟i)+(θi−zk−1)′​uk−1\displaystyle=\arg\min_{\theta_{i}}L^{BC}(\theta_{i};\mathcal{D}_{i})+(\theta_{i}-z^{k-1})^{\prime}u^{k-1} (9)
+ρ2​‖θi−zk−1‖2\displaystyle+\tfrac{\rho}{2}\|\theta_{i}-z^{k-1}\|^{2}
uk\displaystyle u^{k} =uk−1+ρ⁡(θik−zk−1)\displaystyle=u^{k-1}+\rho\,(\theta_{i}^{k}-z^{k-1})
zk\displaystyle z^{k} =(Ωi+Ω¯i−1)−1​(Ωi​θik+Ω¯i−1​θi−1∗).\displaystyle=(\Omega_{i}+\bar{\Omega}_{i-1})^{-1}(\Omega_{i}\theta_{i}^{k}+\bar{\Omega}_{i-1}\theta^{*}_{i-1}).
Ω¯i:=(m+i−1)​Ω¯i−1+Ωim+i,\bar{\Omega}_{i}:=\dfrac{(m+i-1)\,\bar{\Omega}_{i-1}+\Omega_{i}}{m+i}, (10)
Algorithm 1 SAMBAR
Complete original Algorithm 1 SAMBAR
Selectable algorithm with original mathematics
Require: Pretrained VLA πθpre\pi^{\mathrm{pre}}_{\theta} with parameters θ0∗\theta^{*}_{0} (mm pretraining tasks); pretraining importance matrix Ω¯0\bar{\Omega}_{0}; demonstration sets {𝒟i}\{\mathcal{D}_{i}\} for task i∈{1,..,n}i\in\{1,..,n\}; constraint penalty hyperparameter ρ>0\rho>0; total steps KK; learning rate η\eta; number of minibatches NBN_{B}; stabilizer ϵ\epsilon
Ensure: πθn\pi^{n}_{\theta} has high success rate on all m+nm+n tasks
1: for i=1i=1 to nn do
2:       θ←θi−1∗\theta\leftarrow\theta^{*}_{i-1};  u←0u\leftarrow 0
3:       Ωi←Importance​(θi−1∗,𝒟i)\Omega_{i}\leftarrow\textsc{Importance}(\theta^{*}_{i-1},\mathcal{D}_{i})
4:       z←Ω¯i−1⊙θi−1∗+(Ωi+ϵ)⊙θΩ¯i−1+Ωi+ϵz\leftarrow\dfrac{\bar{\Omega}_{i-1}\odot\theta^{*}_{i-1}+(\Omega_{i}+\epsilon)\odot\theta}{\bar{\Omega}_{i-1}+\Omega_{i}+\epsilon}
5:       for k=1k=1 to KK do
6:            sample minibatch b∼𝒟ib\sim\mathcal{D}_{i}
7:            g←∇θLB​C​(θ,b)+u+ρ⁡(θ−z)g\leftarrow\nabla_{\theta}L^{BC}(\theta;b)+u+\rho\,(\theta-z) ⊳\triangleright ∇θℒa\nabla_{\theta}\mathcal{L}_{a}
8:            θ←OptimStep​(θ,g,η)\theta\leftarrow\textsc{OptimStep}(\theta,g,\eta) ⊳\triangleright AdamW [35]
9:            u←u+ρ⁡(θ−z)u\leftarrow u+\rho\,(\theta-z) ⊳\triangleright dual ascent
10:            z←Ω¯i−1⊙θi−1∗+(Ωi+ϵ)⊙θΩ¯i−1+Ωi+ϵz\leftarrow\dfrac{\bar{\Omega}_{i-1}\odot\theta^{*}_{i-1}+(\Omega_{i}+\epsilon)\odot\theta}{\bar{\Omega}_{i-1}+\Omega_{i}+\epsilon}
11:       end for
12:       θi∗←θ\theta^{*}_{i}\leftarrow\theta
13:       Ω¯i←(m+i−1)​Ω¯i−1+Ωim+i\bar{\Omega}_{i}\leftarrow\dfrac{(m+i-1)\,\bar{\Omega}_{i-1}+\Omega_{i}}{m+i}
14: end for
15: return πθn\pi^{n}_{\theta} with θn∗\theta^{*}_{n}
16: function Importance(θ,𝒟\theta,\mathcal{D})
17:       Ω←0\Omega\leftarrow 0
18:       for j=1j=1 to NBN_{B} do
19:            sample minibatch b∼𝒟b\sim\mathcal{D}
20:            Ω←Ω+(∇θLB​C​(θ,b))⊙2\Omega\leftarrow\Omega+\big(\nabla_{\theta}L^{BC}(\theta;b)\big)^{\odot 2} ⊳\triangleright elementwise square
21:       end for
22:       return Ω/NB\Omega/N_{B}
23: end function

VI Simulation Results

The base model isπ0.5. Because it initially performs poorly in LIBERO, it is first pretrained within the benchmark; continual adaptation is then tested from those constructed policies.

Single-task experiments isolate one adaptation, followed by a three-task sequence that tests accumulating skills [13].

VI-A Experimental setup

The broad regime pretrains on 117 tasks and separately adapts to pot-on-stove, items-into-basket or mugs-on-plates. The narrow regime uses 32 tasks, eight per suite, and adapts to pot-on-stove or bowl-in-drawer.

Sequential training starts from 32 tasks and learns pot-on-stove, then bowl-in-drawer, then bottle-on-rack. The final policy is evaluated on all 35 tasks. The experiments use one fixed sequence rather than testing arbitrary task orders.

Complete original Fig. 2: The three finetuning LIBERO tasks we learn in sequence. 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 left to right. Each panel shows the initial scene the policy is given. The first two are long-horizon tasks, while the third is a goal task.
Fig. 2: The three finetuning LIBERO tasks we learn in sequence. 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 left to right. Each panel shows the initial scene the policy is given. The first two are long-horizon tasks, while the third is a goal task.

VI-B Evaluation Metric

Pretrain SR measures old-task performance; Finetune SR measures acquisition. Weighted Avg SR averages by task count, so 117 old tasks dominate one new task in the broad setting. A high weighted score can coexist with poor new-task learning.

Simulation reports five seeds, with five episodes per seed in the broad setting and ten in the narrow/sequential setting. The paper reports 95% confidence intervals. Hardware uses successful rollout counts out of five per task.

VI-C Baselines

Replay-free supervised baselines include SFT, RETAIN, LoRA, EWC and uniform L2 anchoring. A SAMBAR ablation removes importance weighting. Simple Recipe Works uses reinforcement-learning interaction and a short demonstration warm-up; co-training has access to old demonstrations.

Those additional methods use different information and optimization resources. Co-training is a useful replay reference, rather than a proven upper bound on achievable performance [25].

VI-D Training details

Broad adaptation uses batch 64, learning rate 2.5×10⁻⁵ and 500 or 1,000 iterations depending on task. Narrow/sequential baselines use batch 16, learning rate 10⁻⁵ and 10,000 iterations per task; SAMBAR receives 20,000 to let dual updates develop.

The penalty coefficient is 0.02, and importance is estimated from 50 minibatches at each stage’s start. The RL baseline additionally uses a 100-iteration LoRA behavioral-cloning warm-up before environment training.

Results use each method’s best-performing checkpoint. The selection criterion is not fully specified. Comparisons include different training budgets and do not provide a matched wall-clock or memory-cost benchmark.

VI-E Learning from a broadly pretrained policy

For the broadly pretrained policy, forgetting is relatively mild. SAMBAR gives the highest reported weighted mean on pot-on-stove and items-into-basket, with 91.9% and 94.3% old-task success and 92% and 100% new-task success, respectively.

LoRA is stronger on mugs-on-plates:91.9% old-task and 100% new-task success, versus SAMBAR’s 90.7% and 92%. The results do not establish uniform superiority across all tasks.

Complete original TABLE I: Learning one new task from a broadly pretrained policy. We finetune a pi0.5 policy pretrained on 117 LIBERO tasks on finetune tasks, and evaluate it on the 117 pretraining tasks ( Pretrain SR ) and on the finetune task ( Finetune SR ). All entries are success rates in percent, averaged over 5 random seeds with 5 episodes per seed, reported as mean ± \pm 95% confidence interval; higher is better ( ↑ \uparrow ) everywhere. Weighted Avg SR is the task-count-weighted mean over all 118 tasks. Forgetting is mild in this regime, and SAMBAR learns each new task without significantly forgetting the 117 pretrained tasks.
TABLE I: Learning one new task from a broadly pretrained policy. We finetune a pi0.5 policy pretrained on 117 LIBERO tasks on finetune tasks, and evaluate it on the 117 pretraining tasks (Pretrain SR) and on the finetune task (Finetune SR). All entries are success rates in percent, averaged over 5 random seeds with 5 episodes per seed, reported as mean ±\pm 95% confidence interval; higher is better (↑\uparrow) everywhere. Weighted Avg SR is the task-count-weighted mean over all 118 tasks. Forgetting is mild in this regime, and SAMBAR learns each new task without significantly forgetting the 117 pretrained tasks.
Accessible table and original citation links
Method Pretrain
SR ↑\uparrow
Finetune
SR ↑\uparrow
Weighted
Avg SR ↑\uparrow
Finetune task: pot-on-stove
Pretrained (no finetuning) 93.2 ±\pm1.2 4.0 ±\pm11.1 -
SFT 89.0 ±\pm0.9 100.0 ±\pm0.0 89.1
RETAIN [31] 89.7 ±\pm1.2 100.0 ±\pm0.0 89.8
LoRA [19] 88.1 ±\pm0.6 100.0 ±\pm0.0 88.2
EWC [5] 89.5 ±\pm1.4 100.0 ±\pm 0.0 89.6
L2 84.8 ±\pm0.7 84.0 ±\pm11.1 84.8
SAMBAR w/o importance 89.7 ±\pm1.4 56.0 ±\pm32.4 89.5
SAMBAR (ours) 91.9 ±\pm0.9 92.0 ±\pm13.6 91.9
Finetune task: items-into-basket
Pretrained (no finetuning) 93.2 ±\pm1.2 0.0 ±\pm0.0 -
SFT 93.2 ±\pm1.5 100.0 ±\pm0.0 93.3
RETAIN [31] 93.3 ±\pm2.1 96.0 ±\pm11.1 93.4
LoRA [19] 93.4 ±\pm0.9 100.0 ±\pm0.0 93.5
EWC [5] 93.5 ±\pm1.2 100.0 ±\pm0.0 93.6
L2 93.8 ±\pm1.2 92.0 ±\pm13.6 93.8
SAMBAR w/o importance 94.3 ±\pm0.5 36.0 ±\pm11.1 93.8
SAMBAR (ours) 94.3 ±\pm1.3 100.0 ±\pm0.0 94.4
Finetune task: mugs-on-plates
Pretrained (no finetuning) 93.2 ±\pm1.2 0.0 ±\pm0.0 -
SFT 80.7 ±\pm2.0 96.0 ±\pm11.1 80.8
RETAIN [31] 85.4 ±\pm1.1 88.0 ±\pm22.2 85.5
LoRA [19] 91.9 ±\pm1.3 100.0 ±\pm0.0 91.9
EWC [5] 78.8 ±\pm0.8 96.0 ±\pm11.1 78.9
L2 86.1 ±\pm0.6 80.0 ±\pm30.4 86.0
SAMBAR w/o importance 93.2 ±\pm1.2 76.0 ±\pm20.8 93.1
SAMBAR (ours) 90.7 ±\pm1.1 92.0 ±\pm13.6 90.8

VI-F Learning from a narrowly pretrained policy

With narrow pretraining, SAMBAR retains 89.7/91.8% old-task success and learns pot-on-stove/bowl-in-drawer at 86/96%. Uniform SAMBAR retains 91.8/90.6% but acquires those tasks at only 4/38%, illustrating over-anchoring.

The source’s statement that every other baseline falls below 80% old-task success has an exception: the RL baseline retains 84% on pot-on-stove. Replay co-training also retains more than 93% on both comparisons, using historical demonstrations.

Complete original TABLE II: Learning one new task from a narrowly pretrained policy. We finetune a pi0.5 policy pretrained on 32 LIBERO tasks (8 from each suite) on two LIBERO-Long finetune tasks, and evaluate it on the 32 pretraining tasks ( Pretrain SR ) and on the finetune task ( Finetune SR ). All entries are success rates in percent, averaged over 5 random seeds with 10 episodes per seed, reported as mean ± \pm 95% confidence interval; higher is better ( ↑ \uparrow ) everywhere. Weighted Avg SR is the task-count-weighted mean over all 33 tasks. Forgetting is severe in this regime: every baseline forgets the pretraining tasks, whereas SAMBAR remembers the pretraining tasks.
TABLE II: Learning one new task from a narrowly pretrained policy. We finetune a pi0.5 policy pretrained on 32 LIBERO tasks (8 from each suite) on two LIBERO-Long finetune tasks, and evaluate it on the 32 pretraining tasks (Pretrain SR) and on the finetune task (Finetune SR). All entries are success rates in percent, averaged over 5 random seeds with 10 episodes per seed, reported as mean ±\pm 95% confidence interval; higher is better (↑\uparrow) everywhere. Weighted Avg SR is the task-count-weighted mean over all 33 tasks. Forgetting is severe in this regime: every baseline forgets the pretraining tasks, whereas SAMBAR remembers the pretraining tasks.
Accessible table and original citation links
Method Pretrain
SR ↑\uparrow
Finetune
SR ↑\uparrow
Weighted
Avg SR ↑\uparrow
Finetune task: pot-on-stove
Pretrained (no finetuning) 93.9 ±\pm1.4 6.0 ±\pm16.7 -
Co-Training 93.8±\pm1.3 94.0±\pm6.8 93.8
SFT 56.1 ±\pm2.6 96.0 ±\pm6.8 57.3
RETAIN [31] 66.5 ±\pm2.1 94.0 ±\pm11.1 67.3
LoRA [19] 75.4 ±\pm1.8 100.0 ±\pm0.0 76.2
EWC [5] 47.1 ±\pm1.4 96.0 ±\pm11.1 48.6
Simple Recipe Works [25] 84.0±\pm1.3 86.0±\pm16.7 84.1
L2 71.5 ±\pm2.2 74.0 ±\pm16.7 71.6
SAMBAR w/o importance 91.8 ±\pm1.4 4.0 ±\pm6.8 89.1
SAMBAR (ours) 89.7 ±\pm1.7 86.0 ±\pm18.8 89.6
Finetune task: bowl-in-drawer
Pretrained (no finetuning) 93.9 ±\pm1.4 0.0 ±\pm0.0 -
Co-Training 93.1 ±\pm1.7 100.0±\pm0.0 93.3
SFT 23.1 ±\pm1.4 100.0 ±\pm0.0 25.4
RETAIN [31] 27.4 ±\pm1.8 96.0 ±\pm6.8 29.5
LoRA [19] 74.0 ±\pm3.0 100.0 ±\pm0.0 74.8
EWC [5] 19.6 ±\pm0.9 90.0 ±\pm12.4 21.7
Simple Recipe Works [25] 74.1±\pm1.2 92.0±\pm10.4 74.6
L2 64.9 ±\pm2.3 78.0 ±\pm16.2 65.3
SAMBAR w/o importance 90.6 ±\pm1.9 38.0 ±\pm30.9 89.0
SAMBAR (ours) 91.8 ±\pm1.2 96.0 ±\pm6.8 91.9

VI-G Sequential Continual Learning

After all three new tasks, SAMBAR scores 88.2% on old tasks and 84%,70%,82% on the learned tasks, for 87.4% weighted success. Every listed replay-free baseline scores zero on the first new task.

Replay co-training retains 100% on that first task and reaches 95.4% weighted success. Thus the paper’s broad “every baseline forgets” claim should be scoped to the replay-free comparison. SAMBAR still shows some old-task degradation relative to its starting policy.

Complete original TABLE III: Sequential Continual Learning. Starting from the policy pretrained on 32 tasks, we finetune on pot-on-stove , then bowl-in-drawer , then bottle-on-rack . We evaluate the policy after learning all three tasks. All entries are success rates in percent, averaged over 5 random seeds, 10 episodes/seed and reported as mean ± \pm 95% confidence interval; higher is better ( ↑ \uparrow ) for every column. Weighted Avg SR is the task-count-weighted mean over all 35 tasks. After three tasks every baseline has completely forgotten pot-on-stove task, while SAMBAR still performs it 84.0 % 84.0\% of the time along with high success rate on 32 pretrained tasks.
TABLE III: Sequential Continual Learning. Starting from the policy pretrained on 32 tasks, we finetune on pot-on-stove, then bowl-in-drawer, then bottle-on-rack. We evaluate the policy after learning all three tasks. All entries are success rates in percent, averaged over 5 random seeds, 10 episodes/seed and reported as mean ±\pm 95% confidence interval; higher is better (↑\uparrow) for every column. Weighted Avg SR is the task-count-weighted mean over all 35 tasks. After three tasks every baseline has completely forgotten pot-on-stove task, while SAMBAR still performs it 84.0%84.0\% of the time along with high success rate on 32 pretrained tasks.
Accessible table and original citation links
Success on each newly learned task
Method Pretrain
SR ↑\uparrow
pot-on-stove ↑\uparrow bowl-in-drawer ↑\uparrow bottle-on-rack ↑\uparrow Weighted
Avg SR ↑\uparrow
Co-Training 95.4±\pm1.0 100.0±\pm0.0 98.0±\pm5.6 88.0±\pm13.6 95.4
SFT 12.9 ±\pm1.9 0.0 ±\pm0.0 60.0 ±\pm21.5 84.0 ±\pm11.1 15.9
RETAIN [31] 17.4 ±\pm0.7 0.0 ±\pm0.0 68.0 ±\pm16.2 96.0 ±\pm6.8 20.6
LoRA [19] 45.4 ±\pm1.3 0.0 ±\pm0.0 18.0 ±\pm13.6 92.0 ±\pm10.4 44.6
EWC [5] 0.0 ±\pm0.0 0.0 ±\pm0.0 0.0 ±\pm0.0 100.0 ±\pm0.0 2.9
Simple Recipe Works [25] 44.4±\pm3.2 0.0±\pm0.0 36.0±\pm14.2 90.0±\pm8.8 43.8
L2 58.3 ±\pm2.6 0.0 ±\pm0.0 42.0 ±\pm34.5 80.0 ±\pm24.8 56.7
SAMBAR w/o importance 89.8 ±\pm0.8 0.0 ±\pm0.0 0.0 ±\pm0.0 4.0 ±\pm6.8 82.2
SAMBAR (ours) 88.2 ±\pm2.3 84.0 ±\pm6.8 70.0 ±\pm15.2 82.0 ±\pm16.2 87.4

VI-H Ablation Studies

The uniform dual-anchor ablation improves retention over fixed L2 but reduces acquisition. Adding importance to fixed L2 in EWC produces the opposite failure in these experiments: new-task learning remains strong while retention degrades.

Combining importance with dual updates recovers acquisition with modest retention changes relative to uniform SAMBAR. This supports the two components empirically, while the longer SAMBAR schedule remains part of the experimental comparison.

VII Hardware Results

Hardware experiments test four original tabletop tasks and two adaptation tasks on an xArm7. The full tables retain all rollout counts, including methods that preserve old skills but fail the new long-horizon task.

Complete original TABLE IV: Learning one new task on hardware. We pretrain pi0.5 on four tasks and then finetune it on each of the two finetuning tasks. We evaluate every task with 5 rollouts and report the individual tasks as the number of successful rollouts out of 5. Pretrain SR is the percentage of the 20 pretraining rollouts that succeed, Weighted Avg SR is the percentage of all 25 rollouts that succeed. Higher is better ( ↑ \uparrow ) for every column. SAMBAR is the only method that keeps all four pretraining tasks intact while learning the new one.
TABLE IV: Learning one new task on hardware. We pretrain pi0.5 on four tasks and then finetune it on each of the two finetuning tasks. We evaluate every task with 5 rollouts and report the individual tasks as the number of successful rollouts out of 5. Pretrain SR is the percentage of the 20 pretraining rollouts that succeed, Weighted Avg SR is the percentage of all 25 rollouts that succeed. Higher is better (↑\uparrow) for every column. SAMBAR is the only method that keeps all four pretraining tasks intact while learning the new one.
Accessible table and original citation links
Pretraining tasks
Method bird-in-bin ↑\uparrow two-birds-in-bin ↑\uparrow eraser-in-mug ↑\uparrow eraser-in-bowl ↑\uparrow Pretrain
SR ↑\uparrow
Finetune
SR ↑\uparrow
Weighted
Avg SR ↑\uparrow
Finetune task: cake-on-cake-tray
SFT 4/5 1/5 5/5 2/5 60.0 5/5 68.0
RETAIN (α=0.5\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 (α=0.5\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
Complete original TABLE V: Sequential Continual Learning on hardware. Starting from the policy pretrained on the four tasks, we finetune on the cake-on-cake-tray task and then on the remove-lid-and-place-bird task, and evaluate all six tasks after learning the remove-lid-and-place-bird task. Entries follow the same convention as Table IV , with Weighted Avg SR now the percentage of all 30 rollouts that succeed.
TABLE V: Sequential Continual Learning on hardware. Starting from the policy pretrained on the four tasks, we finetune on the cake-on-cake-tray task and then on the remove-lid-and-place-bird task, and evaluate all six tasks after learning the remove-lid-and-place-bird task. Entries follow the same convention as Table IV, with Weighted Avg SR now the percentage of all 30 rollouts that succeed.
Accessible table and original citation links
Pretraining tasks Finetune tasks
Method bird-in-
bin ↑\uparrow
two-birds-
in-bin ↑\uparrow
eraser-in-
mug ↑\uparrow
eraser-in-
bowl ↑\uparrow
Pretrain
SR ↑\uparrow
cake-on-
cake-tray ↑\uparrow
remove-lid-and-
place-bird ↑\uparrow
Weighted
Avg SR ↑\uparrow
SFT 5/5 5/5 5/5 3/5 90.0 4/5 3/5 83.3
RETAIN (α=0.5\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

VII-A Experiment Setup

The authors collect 50 teleoperated demonstrations for each of six tasks. Four train the base policy; cake placement and removing a lid before placing a bird are held out for adaptation.

Hardware training uses batch 18 and learning rate 10⁻⁵. Single-task runs use 10,000 iterations; sequential SAMBAR receives 20,000 per task while baselines receive 10,000. Each task is tested with five complete-success rollouts, without partial credit. RETAIN uses α=0.5 here, versus 0.9 in simulation.

Complete original Fig. 3: The hardware experiment tasks. We pretrain pi0.5 on the four tasks on the left and finetune on the two on the right. Each panel shows the initial scene the policy is given and the arrows shows the task policy needs to perform.
Fig. 3: The hardware experiment tasks. We pretrain pi0.5 on the four tasks on the left and finetune on the two on the right. Each panel shows the initial scene the policy is given and the arrows shows the task policy needs to perform.

VII-B Finetuning on a single task

For cake placement, SAMBAR succeeds in all 25 evaluated old/new-task rollouts. For lid removal and bird placement, it succeeds on all 20 old-task trials and four of five new-task trials, or 24/25 overall.

SFT learns the new task while losing some old performance. RETAIN and EWC retain many old successes but score 0/5 on the new lid task. These small rollout sets illustrate the trade-off without establishing broad hardware reliability.

VII-C Sequential Continual Learning

After learning both hardware tasks in sequence, SAMBAR succeeds on 20/20 old-task trials and 4/5 for each new task, totaling 28/30 or 93.3%. EWC also retains 20/20 old-task successes but achieves 4/5 and 2/5 on the new tasks, totaling 86.7%.

The advantage concerns the combined retention/acquisition balance. Hardware retention alone is not unique to SAMBAR in this final comparison, and five trials per task leave substantial uncertainty.

VIII CONCLUSIONS and FUTURE WORK

The main contribution is a practical replay-free anchoring procedure combining compressed importance statistics and signed dual updates. Its evidence is strongest in the tested narrow-pretraining and sequential settings.

Dependencies include initial old-task importance metadata, checkpoint selection and extra optimization steps. Learning from the robot’s own rollouts is proposed future work.

Original project demonstrations

The authors’ project page provides 19 unique overview, hardware and simulation videos. All original files are retained unchanged with local preview frames. The page reuses the lid-task clips in two comparison galleries; they are archived once each.

The public code repository remains linked. Selected rollouts illustrate behavior alongside the complete quantitative tables.

Project overview
Hardware · SFT · Eraser into bowl
Hardware · SAMBAR · Eraser into bowl
Hardware · SFT · Cake onto tray
Hardware · SAMBAR · Cake onto tray
Hardware · SFT · Remove lid and place bird
Hardware · SAMBAR · Remove lid and place bird
Hardware · SFT · Two birds into bin
Hardware · SAMBAR · Two birds into bin
Hardware · SFT · Cake onto tray
Hardware · SAMBAR · Cake onto tray
Hardware · SFT · Eraser into mug
Hardware · SAMBAR · Eraser into mug
Simulation · SAMBAR · Pot onto stove
Simulation · SAMBAR · Bowl into drawer
Simulation · SAMBAR · Bottle onto rack
Simulation · SFT · Pot onto stove
Simulation · SFT · Bowl into drawer
Simulation · SFT · Bottle onto rack
Concept demonstration

Open image ↗