Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation

Zeva framework overview — **Zeva** performs In-Context Causal Learning (ICCL): a frozen policy learns from the robot’s own physical interactions by extracting action–state-change causality into a dual-timescale causal memory, and injecting it back as a Causal Prompt at decision time. During deployment, all model parameters remain frozen; only the interaction memories are updated.

76.8%

Avg. success on RoboCasa365-Atomic5
(+4.4 pt over Fast-WAM)

83.3%

Avg. success on ChemLab-Evo Level-1 real robot tasks

100%

Pick Up Test Tube SR on ChemLab-Evo
(20 randomized episodes)

Abstract

Generalizable embodied manipulation remains difficult to achieve through pretraining alone, due to unseen physical conditions in the real world. We argue that robots need to learn from their own physical interactions on the fly during real-world deployment and use this knowledge to inform subsequent actions. We present Zeva, the first framework that enables in-context learning from a robot's own physical interaction experience while keeping the policy model frozen. Zeva employs a Causal Interaction Extractor to encode an executed action and its induced state change into a causal interaction signal, which is stored in a dual-timescale causal memory. For subsequent actions, relevant causal interaction signals are retrieved from memory and injected into the frozen policy model as context. Experiments in simulation and real-world manipulation demonstrate that Zeva achieves the best performance among the compared frontier VLAs and WAMs and, more importantly, enables self-evolution during deployment without gradient updates. Its success rate continues to improve as the robot accumulates interaction experience. Furthermore, the acquired interaction experience can generalize across tasks.

Demonstration Videos

The same frozen policy improves across repeated attempts and can be warm-started by a single human demonstration. The rollout-only video is not included here.

Online Self-Evolution 1

Repeated attempts on a real manipulation task: earlier failures are corrected in later attempts using Persistent Interaction Memory, while the policy weights remain frozen.

Online Self-Evolution 2

A second example of gradient-free self-evolution across attempts, showing how causal interaction evidence changes the subsequent execution.

One-Shot Human Warm-Up

A human physically guides the robot once; the same frozen policy later reproduces the demonstrated sequence without any weight update.

Method

Zeva is built around one idea: the robot should not only execute from pretrained weights; it should remember what its own actions caused, and let that memory guide the next attempt.

Stage 1

Causal Interaction Extraction

A Causal Transition Encoder (CTE) integrates visual latents, action encodings, and observed effects into a Causal Interaction State, then projects it into a task Phase Token and a Causal Interaction Signal.

Stage 2

Dual-timescale Causal Memory

A Brief Interaction Trace (BIT) captures recent within-attempt dynamics; a Persistent Interaction Memory (PIM) consolidates cross-attempt evidence within the same episode through similarity-based merging.

Stage 3

In-Context Policy Injection

Zeva retrieves phase-matched interaction evidence and constructs a Causal Prompt for the frozen foundation policy. During deployment, all parameters remain frozen; only the memories are updated online.

Post-Deployment Capability Scaling

Zeva improves as it interacts. With all policy parameters frozen, cumulative success rises across repeated attempts on both simulation and real hardware.

RoboCasa365 cumulative success rate — **RoboCasa365.** Pooled cumulative success rate over five tasks rises from 26% at Evolve 1 to 73% at Evolve 4.

Real-world ChemLab-Evo scaling — **ChemLab-Evo (real ARX manipulator).** Pick Up Test Tube rises from 65% to 100%, Place Beaker from 25% to 70%, Pour Water from 30% to 80%.

Benchmark Results

Method	ElectricKettle	ToasterOvenDoor	Microwave	CoffeeSetupMug	StandMixerHead	Avg.
LingBot-VA	63	60	70	35	48	55.2
Xiaomi-Robotics-1	17	16	82	6	35	31.2
τ₀-WM	15	32	63	12	23	29.0
Fast-WAM	88	66	60	54	94	72.4
Cosmos3-Nano	76	74	76	6	92	64.8
Zeva	78	86	84	44	92	76.8

Success rates (%) on RoboCasa365-Atomic5. Bold marks the best per task; shaded row is Zeva.

Real-World ChemLab-Evo Results

Method	Pick Up Test Tube	Place Beaker	Pour Water	Level 1 Avg.	Titration	Prepare Salt Solution	Level 2 Avg.	Balance Weighing	Extraction	Level 3 Avg.
LingBot-VA	80	70	65	71.7	55	75	65.0	5	0	2.5
Fast-WAM	85	70	75	76.7	45	55	50.0	0	0	0.0
π_0.5	95	60	65	73.3	60	50	55.0	0	0	0.0
Cosmos3-Nano	80	70	60	70.0	45	60	52.5	0	0	0.0
Zeva	100	70	80	83.3	70	70	70.0	5	10	7.5

Success rates (%) over 20 randomized episodes on the real-world ChemLab-Evo benchmark.

Long-Horizon Process Scores

Method	Balance Weighing	Extraction	Avg.
LingBot-VA	28.75	40.00	34.38
Fast-WAM	22.50	58.57	40.54
π_0.5	33.75	31.43	32.59
Cosmos3-Nano	38.75	50.71	44.73
Zeva	47.50	67.14	57.32

Process scores report the average percentage of ordered stages completed on the two long-horizon tasks.

Ablation: Two Memory Timescales

BIT	PIM	Pick Up Test Tube	Place Beaker	Pour Water	Titration	Salt Solution
✓	✓	100	70	80	70	70
	✓	80	50	50	40	55
✓		85	60	70	60	50
		65	40	45	25	35

Real-world ablation (SR %) on five ChemLab-Evo tasks over 20 randomized episodes. Removing PIM reduces SR by 10–20 points; removing BIT causes a larger 15–30 point drop.

What Interaction Memory Changes in the Real World

On the ARX manipulator, red frames are terminal observations from distinct failed attempts; green frames show the subsequent successful attempt. For Pick Up Test Tube and Place Beaker, the two green frames are successive stages of the same continuous successful attempt.

Pick Up Test Tube

Failed attempt

Failed attempt

Successful attempt

Same success (later stage)

Place Beaker

Failed attempt

Successful attempt

Same success (later stage)

Pour Water

Failed attempt

Successful attempt

Cross-Task Effect Retrieval

The VBD effect token captures action-induced state changes beyond task-specific appearance. For each query transition, nearest neighbors are retrieved from other tasks by effect-token cosine similarity. Despite changes in objects, viewpoints, and instructions, the retrieval aligns functionally equivalent effects: container tilting, gripper closure, and post-grasp lifting.

Pouring / Container Tilt

→

→

→

Gripper Closure

→

Lifting / Upward Motion

→

Each block is a before–after transition pair. Effect labels are shown for interpretation and do not enter the similarity ranking.

One-Shot Human Warm-Up

A single human-guided demonstration on Balance Weighing initializes PIM before autonomous execution. The frozen policy then reproduces the demonstrated approach, grasp, transfer, and placement — no gradient update needed. No ICL denotes execution without memory warm-up; ICL uses the demonstrated experience.

Overlaid trajectories — 3D comparison of human (GT), No ICL, and ICL end-effector trajectories after one-shot memory warm-up.

Human warm-up scaling — **Quantitative warm-up effect.** One human demonstration improves or matches SR at every evolution milestone on the three atomic ChemLab-Evo tasks; gains reach 20 points on Place Beaker and 15 points on Pour Water.

BibTeX

@misc{zeva2026,
  title={Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation},
  author={Fu Chen, Xin Ding, Bingjia Huang, Xiangyu Li, Mingju Wang, Jiawei He, Kun Li, Wei Sun, Yunxin Liu, Hao Wu, Ting Cao},
  howpublished={\url{https://air-embodied-brain.github.io/Zeva}},
  year={2026},
}