Vision–Language–Action
Learning generalist robot policies that connect visual understanding, language instructions, and closed-loop action.
Robotics MSc · National University of Singapore
Building robust intelligence for robots that see, predict, and act.
I’m Jiacheng Zhang, a researcher working at the intersection of embodied AI, vision–language–action models, latent world models, and robot learning.
4.44 / 5.0NUS MSc GPA
Top 3Undergraduate cohort
Dean’s ListAcademic distinction
About
Research across academia and industry.
I am pursuing an MSc in Robotics at the National University of Singapore, following a First Class BSc (Hons) in Computer Science (AI) from the University of Liverpool in the United Kingdom, where I graduated in the top three of my cohort.
Across academic labs and industry teams, I have built end-to-end training and evaluation pipelines, developed model-based control algorithms, and tested robot learning systems under systematic distribution shifts.
During my research internship at JD.com, I work on multi-view object grounding, persistent stage intent and foresight-guided residual learning through ObjectBinder-VLA, PSI-WAM and IA-World.
A+ Research ProjectNUS MSc Robotics
First Class HonoursBSc (Hons), CS & AIUniversity of Liverpool · UK
Singapore
Research system
One agenda, three interacting layers.
Hover to hold the three-dimensional orbit.
Latent World Models
Building predictive latent representations for long-horizon planning, fast replanning, and model-based control.
Robust Robot Learning
Improving reliability under camera shifts, sensor noise, visual perturbations, and unseen task conditions.
Selected work
From latent world models to reliable manipulation.
My research at JD.com connects three questions: what should a robot attend to, when should its intent change, and how can predicted action consequences improve control?
01
Multi-view object grounding · JD.com
ObjectBinder-VLA
ICLR 2027 submission · First author
A predictive object-state interface for VLA policies. Permutation-consistent multi-view binding represents object location, visibility and absence; intent-conditioned dynamics predict future object states for action generation. These explicit states augment global visual–language context, while the segmentation teacher is used only during training.
Research origins Earlier LaWAM reproduction and synchronized wrist-view prediction experiments revealed the value of wrist-camera evidence, motivating this explicit multi-view object-state design.
98.10%Original LIBERO · four suites
76.37%LIBERO-Plus · seven-axis mean
Final 50K checkpoint: 1,962 / 2,000 successes on Original LIBERO. LIBERO-Plus reports an unweighted mean across seven perturbation axes, evaluated over 10,030 episodes.
02
Frozen VLA + residual RL
Uncorrected base planPredict Base Future
Foresight-guided residualCorrect, then execute
Foresight-guided residual RL · JD.com
IA-World
ICLR 2027 submission · Fourth author
A latent world model predicts the consequences of a frozen VLA’s uncorrected action plan. This Base Future guides bounded residual corrections before execution, while changes in goal-conditioned foresight scores provide dense progress feedback for online reinforcement learning. The pretrained VLA stays frozen.
72% → 94%LIBERO · Soup & Sauce in Basket
42% → 62%ManiSkill · Peg Insertion
Frozen base → IA-World after 15K environment steps. GR00T on the LIBERO task (100 evaluation episodes); π0.5 on the ManiSkill task (50 episodes). These are task-level results, not benchmark-wide averages.
03
Slow–fast VLA
Stage-level intentPersistent intent
World model + stopping headRetain or refresh
Persistent intent & world-action models · JD.com
PSI-WAM
AAAI 2027 submission · Third author
A slow–fast VLA that keeps stage-level intent fixed across action chunks while replanning from fresh observations. An action-conditioned latent world model forecasts the consequences of the proposed chunk; a learned stopping head predicts stage completion, deciding when to truncate the chunk and refresh intent or retain it for the next request.
98.1%LIBERO · four-suite average
80.2%Seven perturbation families
83.2%RoboTwin 2.0 · six-task protocol
The stopping decision uses predicted stage completion, not a fixed prediction-error threshold. Results refer to each paper-specific evaluation protocol.
04
Multi-agent latent world models · NUS MARMot Lab
Role-Aware MA-TD-MPC
At NUS MARMot Lab, I adapted TD-MPC2 for two tugboats cooperatively controlling a barge in simulation, using a latent world model and CEM planning. I then developed Role-Aware MA-TD-MPC with role-conditioned representations, teammate reconstruction and consensus modeling. This year-long ME5400 research project received an A+.
+13%Final evaluation return vs. MA-TD-MPC
05
LLM agents · Singapore–MIT SMART
RoboAtlas
As a Visiting Researcher at SMART, I led RoboAtlas, an LLM-based system for organising and checking information about robotics companies and technologies. The pipeline combines structured extraction, automated verification and human review. I am the first author of the work accepted to the ICRA 2026 Late Breaking Results track.
ICRA ’26 LBRAccepted · First author
Publications & manuscripts
Research in motion.
Research on object-centric VLA, persistent intent, residual reinforcement learning and robotics landscape intelligence. Submission status is shown separately from accepted work.
01
ObjectBinder-VLA: Permutation-Consistent Multi-View Grounding and Intent-Conditioned Dynamics
2026
02
IA-World: Residual VLA Post-Training with Latent World Model
2026
03
PSI-WAM: Persistent Stage-Aware Intent from World Action Models for Robot Manipulation
2026
04
The RoboAtlas: Mapping the Global Robotics Landscape
2026
05
SMaRT-Tug: Structured Multi-Agent Reinforcement Learning for Physics-Based Tugboat–Barge Collaborative Manipulation
2026