TuKA++: Multi-Factor Lifelong Embodied Navigation Learning with Tucker Adaptation

Xudong Wang*,1,2,3, Gan Li*,1,2, Zhi Han†,1, Chen Gao3,4, Yue Liao3, Dongyue Wang1,2, Zhiyu Liu1,2, Yao Wang5, Lianqing Liu1, Senior Member, IEEE, Si Liu4,6, Senior Member, IEEE, Shuicheng Yan3, Fellow, IEEE
1 State Key Laboratory of Robotics and Intelligent Systems, Shenyang Institute of Automation, Chinese Academy of Sciences
2 University of Chinese Academy of Sciences
3 School of Computing, National University of Singapore
4 School of Artificial Intelligence, Beihang University
5 Center for Intelligent Decision-making and Machine Learning, School of Management, Xi'an Jiaotong University
6 Hangzhou Innovation Institute, Beihang University
* Equal contribution.   Corresponding author.
Shenyang Institute of Automation, CAS · UCAS · NUS · Beihang University · Xi'an Jiaotong University
Illustration of the proposed Multi-Factor Lifelong Embodied Navigation (MFLEN) problem. a) Traditional single-factor lifelong learning, where the agent continually learns along a single task dimension and updates its knowledge sequentially. b) Multi-factor lifelong learning, where tasks evolve along multiple coupled dimensions, requiring the agent to decouple and represent different factor-level knowledge instead of treating each task as an isolated unit. c) In the proposed MFLEN task, a navigation agent continually learns from a sequence of coupled navigation scenarios and decouples scene knowledge, environment knowledge, and instruction-style knowledge. The goal of the proposed MFLEN is to dynamically compose the learned factor-level knowledge during inference and evolve toward a universal embodied navigation agent for all-scenario navigation.

Abstract

Recent large-scale vision-language models have significantly advanced embodied navigation across diverse paradigms. However, existing navigation agents are typically trained or adapted under static task settings, making it difficult to continually acquire new knowledge in dynamic open-world environments without forgetting previously learned scenarios. Moreover, real-world navigation scenarios are usually determined by multiple coupled factors, such as physical scenes, environmental conditions, and instruction styles, while existing LoRA-based adaptation methods mainly learn task-level matrix adapters and cannot explicitly decouple such hierarchical knowledge. To address these limitations, we formulate a new problem, termed Multi-Factor Lifelong Embodied Navigation (MFLEN), where an agent is required to continually learn from coupled navigation scenarios, decouple factor-level knowledge, and dynamically compose learned knowledge during task-agnostic inference. To solve MFLEN, we propose Tucker Adaptation++ (TuKA++), a generalized high-order tensor adaptation framework that represents multi-factor navigation knowledge through Tucker decomposition. It decouples scene, environment, and instruction-style knowledge into factor-specific expert matrices, and aligns the high-order tensor representation with two-dimensional LLM parameter matrices through tensor contraction. We further design a Factor-wise Knowledge Inheritance and Exploration strategy and a progressively expandable shared Tucker subspace with dynamic zero-padding to support continual learning while structurally preserving old task adapters. Extensive experiments demonstrate that TuKA++ consistently outperforms state-of-the-art baselines.

Method

Comparison between LoRA, MoE LoRA, and our TuKA++ for representing coupled navigation knowledge. a) Vanilla LoRA directly learns a task-specific adapter for each coupled navigation scenario, where scene, environment, and instruction-style knowledge are entangled into a single matrix-level representation. b) MoE-based LoRA methods represent a sequence of coupled navigation scenarios with a shared matrix $A$ and multiple matrices $\{B_n\}$. Although this design efficiently explores shared-specific knowledge across tasks, each expert cannot explicitly decouple multi-level knowledge factors. c) TuKA++ represents coupled knowledge in a high-order tensor space. Scene, environment, and instruction-style knowledge are decoupled into independent factor expert matrices, and the corresponding experts can be dynamically composed during inference.
Illustration of the proposed Decoupled Knowledge Incremental Representation for decouple learning the multi-hierarchical navigation knowledge in a high-dimensional tensor space. Our TuKA++ performs decoupled knowledge learning with Factor-wise Knowledge Inheritance and Exploration strategy, and performs anti-forgetfulness learning with Progressively Expandable Shared Tucker Subspace strategy.
For the $t$-th coupled scenario $T_t=(S_s,E_e,L_p)$ (scene $s$, environment $e$, instruction-style $p$), the per-layer adapter is $$\Delta W_t^{l} = U_1^{l}\cdot\big(\mathcal{G}^{l}\times_3 U_3^{l}[s,:]\times_4 U_4^{l}[e,:]\times_5 U_5^{l}[p,:]\big)\cdot (U_2^{l})^{\top},$$ where $\mathcal{G}$ is the shared core tensor, $U_1,U_2$ are the shared encoder/decoder factors, and $U_3,U_4,U_5$ are the scene, environment, and instruction-style expert factors. Old-task adapters are preserved bit-exactly via a progressively expandable shared subspace with dynamic zero-padding.
Illustration of catastrophic forgetting in continual skills learning. The new skill adaptation leads to catastrophic forgetting of old skills.

Algorithms

The TuKA++ training pipeline and the task-agnostic inference procedure.

Algorithm 1: TuKA++ training pipeline.
Algorithm 2: TuKA++ task-agnostic inference.

MFLEN Benchmark

Illustration of the proposed MFLEN benchmark. It contains 40 multi-factor embodied navigation tasks constructed from different combinations of scenes, environmental conditions (normal, low-light, overexposure, and scattering), and instruction styles (Vision-and-Language Navigation (VLN), Object Localization Navigation (OLN), and Dialogue-based Understanding Navigation (DUN)). Each task is coupled with the three types of knowledge. The first 30 tasks are used for sequential learning, the remaining 10 tasks are held out to evaluate generalization to factor combinations.

Allday-Habitat Simulator

Examples of navigation scenarios generated by the proposed Allday-Habitat simulator. The platform extends the standard Habitat simulator from normal visual conditions to diverse environmental degradations, including (a) normal, (b) low-light, (c) overexposure, and (d) scattering.

Real-World Robotic Platform

Illustration of the real-world robotic navigation platforms.

Navigation Demonstrations

Task-wise rollouts across the 40-task MFLEN benchmark, in order. Tasks 1–30 are learned sequentially (lifelong learning); Tasks 31–40 are held-out generalization tasks (inference only, not trained).

Lifelong Learning Tasks 1–30 · continual learning

Task 1
VLN Scattering E9uDoFAP3SH

Walk into the room and turn left. Walk right past the staircase and into the hall to the right of the chair. Walk straight ahead into the large room and stop next to the rug.

Task 2
DUN Normal 2n8kARJN3HM

Find the lamp. Navigator: Im guessing straight thru the doorway? Oracle: Yes, go through the doorway and turn right to the stairs.

Task 3
VLN Low Light E9uDoFAP3SH

Leave the bedroom, and take a right. Go into the hallway straight across. Enter the sitting room. Stop once you reach the rug.

Task 4
DUN Normal b8cTxDM8gDG

Find the book. Navigator: Should I go through the door in front of me? Oracle: Yes go through the door and keep going straight past spider man Navigator: Should I go through the door to the left of the window or continue? Oracle: Go straight then head down the stairs on your left Navigator: Should I go left or right at the bottom of the stairs? Oracle: Go right then go right again to the landing of the next set of stairs.

Task 5
OLN Low Light Real-World 2

Go around the left corner and find the red ball behind the white line.

Task 6
DUN Scattering B6ByNegPMKs

Find the sign. Navigator: Should I go to the right or left hallway? Oracle: Go right and keep going straight down the hallway Navigator: should I keep going down this hallway still? Oracle: Just keep going straight till you get to an intersection.

Task 7
DUN Scattering uNb9QFRL6hY

Find the box. Navigator: Do we go left or right? Oracle: Right, down the stairs and towards the kitchen.

Task 8
VLN Low Light Real-World 5

Follow the cones forward, then turn left at the blue cone and stop in front of the red ball.

Task 9
DUN Low Light B6ByNegPMKs

Find the microwave. Navigator: I'd guess I'm looking for a breakroom. The place must be an nightmare if you're new. Where should I be going? Oracle: Definitely! Move down the hall directly in front of you a bit. Navigator: Keep going or did I miss it? Oracle: Nope, it's showing to keep on a going straight. Navigator: Thought it'd be near those chairs. Guess not. Where to? Oracle: Go straight to the hallway, and take a right and head down it. This one is rough lol.

Task 10
OLN Scattering EDJbREhghzL

Go to the lounge on level 1 and remove all the items from the top of the fireplace.

Task 11
VLN Normal PX4nDJXEHrG

Go to the right of the stairs and into the large master bedroom. Wait at the doorway to the bathroom.

Task 12
DUN Normal Real-World 2

Find the yellow cylinder. Navigator: Should I go straight or turn? Oracle: Go straight past the cone, then turn left at the corner; the yellow cylinder is right next to the wall.

Task 13
OLN Low Light vyrNrziPKCB

Go to the bathroom on level 2 next to the rec room and refold the hand towels to the right of the sink.

Task 14
DUN Overexposure EDJbREhghzL

Find the toiletpaper. Navigator: Should I proceed inside, or take the path towards the pool? Oracle: You should go straight, turn left at the circle, then go right toward the far end of the building.

Task 15
VLN Low Light Real-World 4

Walk straight ahead, pass between the two orange cones, then through the cone-shaped barrels, and stop in front of the red ball.

Task 16
OLN Normal vyrNrziPKCB

Move to the hallway next to the kitchen and turn the knob on the sink.

Task 17
OLN Normal 82sE5b5pLXE

Go to the end of the hallway and bring me the pot next to the doorway.

Task 18
VLN Scattering cV4RVeZvu5T

Head towards the fireplace and pass by it with the fireplace to your right. Stop in front of the sliding glass door.

Task 19
VLN Overexposure 1pXnuDYAj8r

Exit the storage area, then turn left and move through the bedroom into the hallway. Make a sharp left and move three steps down the stairway.

Task 20
DUN Normal Real-World 4

Find the yellow hemisphere. Navigator: Which way should we go now? Oracle: Go forward, pass the yellow and blue cones on the left, then go forward and stop at the yellow hemisphere.

Task 21
OLN Normal EDJbREhghzL

Go to the lounge on level 1 and on the wall above the stand with the lamp remove the bottom of the two vertically adjacent pictures.

Task 22
OLN Overexposure cV4RVeZvu5T

Go to the bathroom on first level and pick up the towel next to the sink.

Task 23
VLN Normal 2n8kARJN3HM

Walk down the two flights of stairs, stop at the end of the stairs.

Task 24
DUN Low Light S9hNv5qa7GM

Find the towel. Navigator: Whoops! I had to guess this room! Where do I head after I go through the laundry room? Oracle: take a left into the laundry room the a right out of the door leading into the kitchen.

Task 25
DUN Normal uNb9QFRL6hY

Find the box. Navigator: where do I go from this room? Oracle: go left past the dining room table and up the steps by the front door Navigator: where do i go from the bathroom? Oracle: go forward then right and go down the first set of steps"

Task 26
OLN Overexposure EDJbREhghzL

Walk into the study and face the fireplace and pick up the trinket on the fireplace shelf.

Task 27
VLN Low Light Real-World 1

Walk forward, passing the small ball to the right, then the yellow cone to the left, and continue walking along the cones until you reach the orange ball, then stop.

Task 28
VLN Low Light r1Q1Z4BcV1o

Move past the table and enter the door with a garbage bin next to it. Stand near the entrance.

Task 29
DUN Scattering S9hNv5qa7GM

Find the nightstand. Navigator: Ok, I see a sort of night stand but I guess I need help. Oracle: Walk around past the dresser, out the door to your right and on to the bottom of the staircase Navigator: Oops I ended up in someone's bedroom. Oracle: Turn around and go out down the hallway until you reach a bathroom.

Task 30
OLN Normal Real-World 1

Past the blue cone and the red ball, find the yellow ball hidden in the grass.

Generalization Tasks 31–40 · held-out, inference only

Task 31
VLN Low Light 82sE5b5pLXE

Go to the bathroom that's connected to the bedroom and reach up and wipe the light with a soft cloth.

Task 32
VLN Scattering kEZ7cmS4wCh

Turn to the right and go down the stairs. At the bottom of the stairs turn left and proceed out to the patio. Stop right there, you will see the pool in front of you. Wait there.

Task 33
DUN Overexposure b8cTxDM8gDG

Find the tv. Navigator: Shall I continue, see for further TV's or head closer? Oracle: Turn to the right and go straight down that hallway until you reach the stairs. Navigator: I unintentionally did not make a movement, just go ahead and repeat your instructions, sorry about that Oracle: Turn to the right and go straight down that hallway until you reach the stairs.

Task 34
OLN Low Light 1pXnuDYAj8r

Navigate to the shower located in the bathroom. The bathroom is accessible from the bedroom, which lies at the end of the hallway near the wall with the horse painting. Stop once you are directly in front of the shower.

Task 35
OLN Overexposure 82sE5b5pLXE

Go to the bathroom on level one next to the master bedroom and turn off the light inside the shower stall.

Task 36
VLN Scattering 1pXnuDYAj8r

Go right past the painting of stones. Continue going past the staircase and turn left at the open doorway on the left. Stop at the laundry machine.

Task 37
OLN Normal r1Q1Z4BcV1o

Go to the main room and the large open central area outside the bathroom.

Task 38
VLN Normal Real-World 3

Walk forward to the blue cone, then pass through the right side after reaching it, go between the cones, and stop in front of the blue cone.

Task 39
DUN Low Light Real-World 3

Navigator: How do I get to the end of the cone path? Oracle: Pass to the right of the red cone, then follow the cones straight ahead. Navigator: Okay, where do I stop? Oracle: Stop in front of the yellow cone at the end.

Task 40
OLN Normal Real-World 5

Walk to the other side of the corridor and follow the cone-shaped barrels to find the final yellow cylinder.

Experimental Results

TuKA++ consistently outperforms state-of-the-art continual-adaptation baselines under the MFLEN setting.

Task-wise Success Rate (SR ↑, %) under the MFLEN setting. Tasks 1–30 are lifelong-learning tasks; Tasks 31–40 are unseen factor-combination generalization tasks.
Task-wise Forgetting Success Rate (F-SR ↓, %) under the MFLEN setting. Tasks 1–30 are lifelong-learning tasks; Tasks 31–40 are unseen factor-combination generalization tasks.
Radar chart comparison of SPL, F-SPL, OSR, and F-OSR under the MFLEN setting.
Comparison between the StreamVLN backbone and TuKA++ on the MFLEN benchmark.