LoongForge-VLDA

Vision-Language-Delta-Action: learning compact Deltas as actionable transition interfaces for robotic manipulation.

Lingsen Zhang^1,2,* Wei Li^1,2,* Jie He^1,2 Henghua Zhang² Kai Mo² Shiyong Li² Yanpeng Wang² Rui Shao^1,3,†

¹ Harbin Institute of Technology, Shenzhen · ² Baidu, Inc. · ³ Shenzhen Loop Area Institute

* Equal contribution † Corresponding author

Paper ↗ Code ↗ Model ↗

98.5%LIBERO average

84.5%LIBERO-Plus, zero-shot

62.9%RoboTwin 2.0, 10 hard tasks

73.0%Real robot · R1 Lite

72.0%Real robot · AgileX

1 GPUReal-world inference on a single GPU

01The idea

Don't predict the whole future. Capture the change.

What does an action actually change? Manipulation is a chain of state transitions, and most of every frame stays the same. VLA models must infer that change implicitly; world-action models predict whole future frames and spend capacity on pixels that never move.

VLDA distills only what changes into a compact Delta and hands it to the action head as an explicit interface: understand the scene, express the transition, then execute.

VLA, WAM and VLDA compared — **Figure 1.** VLA, WAM and VLDA side by side. VLDA keeps the transition signal and drops the redundant future pixels.

UnderstandVision and language are encoded into a task-aware representation.

Transition · ΔA compact Delta captures the local and global change the task requires.

ExecuteThe action head is refined by the Delta, with no future frames needed at inference.

VLA Transition modeling Flow-matching actions

02The method

From a Delta to an action, in three moves.

A V-JEPA 2 EMA teacher provides the target transition. Two filters distill it into a Delta; one refinement module feeds it back into the action head.

01 / LT-FILTER

Local Transition Filter

Selects the top-64 regions with the largest transition magnitude and matches them to the predicted Delta by Sinkhorn optimal transport, so the Delta is not diluted by static background.

02 / GT-ANCHOR

Global Transition Anchor

Anchors the local patches to a global summary of the transition, keeping the Delta consistent with the task as a whole.

03 / DDAR

Delta-Driven Action Refinement

DAMG grounds the Delta on the action manifold with an InfoNCE loss. DSAF seeds the action flow with a Delta-based noise bias and a prefix token.

**Figure 2.** The VLDA architecture: the teacher builds the target transition, LT-Filter and GT-Anchor distill the Delta, and DDAR feeds it into action generation.

STAGE 1

Learn the transition

Pre-train the Delta from the teacher, without actions.

L1 = L_LTF + L_GTA

STAGE 2

Learn to act

Train the action head with the Delta as an interface, keeping the stage-1 loss as a regularizer.

L2 = L_act + λ·L1 + λ_DDAR·L_DDAR

Built on LoongForge. VLDA is trained with the LoongForge embodied stack: DDP, ZeRO-1, FSDP and HSDP, component-wise freezing and checkpointing. Training used 8 GPUs. Inference needs no future observations, and the EMA teacher can be dropped after training.

03The results

Stronger where it counts, consistently.

Three simulation benchmarks and a set of ablations. Our row is highlighted.

98.5%LIBERO average

Top average, with the biggest lift on long-horizon tasks.

VLDA beats π0.5 by 1.6 points on average and edges out every other baseline, reaching 95.2% on the Long suite.

Spatial · Object · Goal · Long

Success rate (%) on LIBERO

Method	Spatial	Object	Goal	Long	Avg
VLDA	99.8	99.6	99.4	95.2	98.5
π0.5	98.8	98.2	98.0	92.4	96.9
VLA-Adapter	97.8	99.2	97.2	95.0	97.3
OpenVLA-OFT	97.6	98.4	97.9	94.5	97.1
CronusVLA	97.3	99.6	96.9	94.0	97.0
InternVLA-M1	98.0	99.0	93.8	92.6	95.9
UniVLA	95.4	98.8	93.6	94.0	95.4
π0	96.8	98.8	95.8	85.2	94.2
GR00T-N1	94.4	97.6	93.0	90.6	93.9

+10.9points over π0.5, zero-shot

Robust when the world is perturbed.

Ahead of π0.5 on all seven perturbation dimensions, and ahead of the strongest baseline, GuidedVLA, on six of them (+9.1 on average). The gap widens on the Long suite: +13.4 over π0.5 and +12.1 over GuidedVLA.

7 perturbation dimensions · 10,030 rollouts

Zero-shot success rate (%) on LIBERO-Plus

Method	Camera	Robot	Language	Light	Background	Noise	Layout	Avg
VLDA	72.4	75.3	84.0	96.0	95.5	89.4	85.7	84.5
GuidedVLA	73.7	51.4	62.6	94.6	89.0	85.2	79.9	75.4
π0.5	59.7	65.5	75.3	87.0	82.4	72.1	80.3	73.6
DreamVLA	65.0	40.8	63.5	85.7	82.6	84.9	74.0	69.9
OpenVLA-OFT	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
π0	13.8	6.0	58.8	85.0	81.4	79.0	68.9	53.6

+20.2points over π0.5, 10 hard tasks

Biggest gains on complex state-transition tasks.

Best or tied-best on every selected task. Put Bottles Dustbin jumps +46, Stacking Three +30, Place Dual Shoes +27 and Open Microwave +27. Across all 50 tasks VLDA scores 50.86% against 43.84% for π0.5.

10 selected hard tasks · 50-task average 50.86% vs 43.84%

Success rate (%) on RoboTwin 2.0

Method	Ranking RGB	Ranking Size	Grab Roller	Open Laptop	Open Microwave	Place Dual Shoes	Put Bottles Dustbin	Shake Bottle	Stacking Three	Stacking Two	Avg
VLDA	53	16	96	60	64	34	55	100	65	86	62.9
π0.5	35	14	89	35	37	7	9	100	35	66	42.7
π0	5	1	80	46	50	0	13	60	24	41	32.0
RDT	0	0	43	32	20	4	4	45	17	30	19.5

+24points on Basket Relocation

Every module earns its place.

Removing any single piece costs accuracy. LT-Filter and GT-Anchor each help alone, and together they help most, notably on the real robot.

Basket Relocation, R1 Lite: 48% without both → 72% with both

Ablations (LIBERO average unless noted)

w/o DSAF

~~98.5~~97.2

Delta-seeded action flow removed

w/o transition pre-training

~~98.5~~97.3

Stage 1 skipped

Full-future prediction

~~98.5~~97.8

Predicting whole frames instead of the Delta

w/o DAMG (LIBERO-Plus)

~~84.5~~82.0

Light +6.3, Background +3.7 with DAMG

LT-Filter and GT-Anchor

LT-Filter	GT-Anchor	LIBERO Spatial	Object	Goal	Long	Ring	Sweep	Sort	Basket
✗	✗	98.6	98.4	98.0	92.8	76	52	48	48
✓	✗	99.4	99.4	99.0	94.4	84	60	56	68
✗	✓	98.8	99.0	98.8	93.8	80	60	52	60
✓	✓	99.8	99.6	99.4	95.2	92	68	60	72

Simulation columns are LIBERO success rate (%); real-world columns are Ring Stacking, Table Sweeping, Conveyor Sorting and Basket Relocation on R1 Lite. Each module helps alone; together they help most.

04In the real world

Still ahead on real robots.

Four long-horizon tasks on a GALAXEA R1 Lite, 25 trials each, plus a second platform, AgileX Cobot Magic. Numbers are final-stage success rates (%).

Ring Stacking

92%vs. π0.5 80%

VLDA

π0.5

OpenVLA-OFT

Table Sweeping

68%vs. π0.5 52%

VLDA

π0.5

OpenVLA-OFT

Conveyor Sorting

60%vs. π0.5 44%

VLDA

π0.5

OpenVLA-OFT

Basket Relocation

72%vs. π0.5 48%

VLDA

π0.5

OpenVLA-OFT

+17.0points over π0.5, R1 Lite

Averaged over four tasks, VLDA reaches 73.0% against 56.0% for π0.5. On a second robot the lead holds. Basket Relocation uses an out-of-distribution mat position, and VLDA still reaches 72% at the final placement.

Platform (average)	OpenVLA-OFT	π0.5	VLDA
GALAXEA R1 Lite	16.0	56.0	73.0
AgileX Cobot Magic	17.0	54.0	72.0
Δ vs. π0.5 (pp)			+17.0 / +18.0

What the Delta looks at

The Delta concentrates on the regions that move and interact, not on the static scene.

Delta visualization — **Figure 3.** Visualization of the learned Delta during manipulation.

05Build with VLDA

Code and citation.

GitHub ↗

LoongForge-VLDA Cookbook

Training code, configs and recipes for VLDA live in LoongForge.

git clone https://github.com/baidu-baige/LoongForge.git

Citation

@misc{zhang2026visionlanguagedeltaactionlearningcompactdeltas,
  title={Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation},
  author={Lingsen Zhang and Wei Li and Jie He and Henghua Zhang and Kai Mo and Shiyong Li and Yanpeng Wang and Rui Shao},
  year={2026},
  eprint={2610.25874},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.25874},
}

Read the preprint on arXiv ↗.