EmbodiSteer

Steering Embodiment-Agnostic Visuomotor Policies with Joint-Space Guidance for Zero-Shot Cross-Embodiment Deployment

Shihefeng Wang^*, Kangchen Lv^*, Mingrui Yu^†, Xiang Li^†

Tsinghua University | Beijing Key Laboratory of Embodied Intelligence Systems | BNRist

^*Equal contribution. ^†Corresponding authors.

CoRL 2026 Spotlight✨

Paper

arXiv BibTeX Code

Overview Video

Sound on

Abstract

Scalable robot imitation learning relies on large-scale heterogeneous data from diverse robots or body-free data, making Cartesian end-effector actions a key interface for embodiment-agnostic policy learning. However, end-effector-only abstraction leaves Cartesian policies unaware of the deployed robot body, making them brittle under robot-specific constraints such as whole-body collision avoidance. To overcome this limitation, we present EmbodiSteer, a training-free framework that steers embodiment-agnostic visuomotor policies toward zero-shot, embodiment-aware deployment. EmbodiSteer keeps policy learning in Cartesian space while efficiently lifting inference-time diffusion sampling into the target robot's joint space via forward kinematics and Jacobian-based updates. With whole-body collision-aware guidance over joint trajectories after each denoising step, the arm can be steered away from collisions while preserving learned end-effector behavior. Across 9 simulated robots, EmbodiSteer reduces collision rate by 46.1 percentage points and increases task success by 28.5 percentage points over Cartesian-only execution; on two physical robots in highly constrained scenarios, the respective improvements are 90.0 and 36.7 percentage points.

+28.5 pp simulation success rate

-46.1 pp simulation collision rate

+36.7 pp real-world success rate

-90.0 pp real-world collision rate

All improvements are absolute changes in percentage points (pp), relative to Cartesian-only execution.

EmbodiSteer teaser showing cross-embodiment deployment with obstacle-aware guidance

Real-World Results

Same UMI-trained checkpoint, zero-shot deployment on two robot embodiments

Compare Base Cartesian Policy and EmbodiSteer rollouts on UR5 and Franka Panda. Each task shares one checkpoint, trained on 200 obstacle-free UMI demonstrations, across the two robots. Obstacles are introduced only at deployment.

All rollouts are shown at 2x speed.

Make Iced Coffee

Obstacle layout

UR5 Base Cartesian Policy

UR5 EmbodiSteer

Franka Panda Base Cartesian Policy

Franka Panda EmbodiSteer

Put Flower in Vase

Obstacle layout

UR5 Base Cartesian Policy

UR5 EmbodiSteer

Franka Panda Base Cartesian Policy

Franka Panda EmbodiSteer

Arrange Banana

Obstacle layout

UR5 Base Cartesian Policy

UR5 EmbodiSteer

Franka Panda Base Cartesian Policy

Franka Panda EmbodiSteer

Real-world results by robot and task

TSR and COR are success and collision counts over 10 trials per robot-task pair. Task success and obstacle contact are evaluated independently. The total aggregates 60 trials, as in Table 12 of the paper.

Robot	Task	No Obstacles, Base Policy	Obstacles, Base Policy		Obstacles, EmbodiSteer
Robot	Task	TSR ↑	TSR ↑	COR ↓	TSR ↑	COR ↓
UR5	Make Iced Coffee	9/10	4/10	10/10	8/10	1/10
	Put Flower in Vase	7/10	4/10	9/10	6/10	1/10
	Arrange Banana	8/10	4/10	9/10	9/10	0/10
Franka Panda	Make Iced Coffee	9/10	6/10	10/10	8/10	1/10
	Put Flower in Vase	6/10	1/10	10/10	6/10	0/10
	Arrange Banana	10/10	6/10	10/10	10/10	1/10
Total	All tasks	49/60	25/60	58/60	47/60	4/60

Simulation Results

Evaluation across 9 robot embodiments

Quantitative results in simulation across 9 robot embodiments

TSR is task success rate (%), RWD is maximum reward per episode, and COR is collision occurrence rate (%). Values are averaged over nine robot embodiments, as in Table 1 of the paper.

Obstacles	Method	PlaceToast			TurnOnFaucet			MakeCoffee			Average
Obstacles	Method	TSR ↑	RWD ↑	COR ↓	TSR ↑	RWD ↑	COR ↓	TSR ↑	RWD ↑	COR ↓	TSR ↑	RWD ↑	COR ↓
w/o Obs.	EE	96.8	.980	--	83.6	.937	--	89.7	.932	--	90.0	.950	--
w/o Obs.	Joint	94.4	.965	--	87.9	.957	--	89.7	.929	--	90.7	.950	--
w/ Obs.	EE	43.4	.634	52.6	41.6	.752	61.8	22.2	.456	58.3	35.7	.614	57.6
	EE w/ Sampling	47.1	.661	51.3	43.9	.767	53.2	24.1	.478	60.0	38.4	.635	54.8
	EE w/ CBF	65.8	.764	15.3	56.3	.797	29.9	48.9	.606	8.6	57.0	.722	17.9
	Joint w/ CG	57.7	.687	0.2	10.3	.403	59.6	22.4	.424	40.6	30.1	.505	33.5
	EmbodiSteer	74.8	.829	4.6	60.6	.833	27.1	57.2	.670	2.8	64.2	.777	11.5

Simulation results across three evaluation seeds

Mean ± standard deviation across three seeds, averaged over the three tasks and nine robot embodiments under obstacles, as in Table 2 of the paper.

Method	TSR (%) ↑	RWD ↑	COR (%) ↓
EE	36.0 ± 0.2	.614 ± .004	56.8 ± 0.9
EE w/ Sampling	37.7 ± 0.5	.621 ± .012	53.1 ± 1.3
EE w/ CBF	57.2 ± 0.6	.726 ± .005	18.3 ± 0.3
Joint w/ CG	30.3 ± 0.5	.503 ± .003	33.5 ± 0.6
EmbodiSteer	63.3 ± 1.4	.771 ± .011	11.7 ± 0.8

Demonstration coverage

Collision guidance perturbs policy-proposed motion, making the coverage of training demonstrations relevant to task performance. On MakeCoffee, we vary the number of demonstrations (20, 50, 100, or 200) and, separately, trajectory diversity. Diversity scales the waypoint randomization used during motion-planning data generation to 10%, 33%, 67%, or 100% of the original setting, with 200 demonstrations at each level. Increasing either quantity or diversity improves task success under guidance.

MakeCoffee success rates for EE, EE w/ CBF, Joint w/ CG, and EmbodiSteer as demonstration count and waypoint randomization increase — Demonstration quantity (left) and trajectory diversity (right), from Figure 6 of the paper. Select the plot to view it at full resolution.

Alternative constraint integration

We also implement a JM2D-style baseline that combines the same frozen Cartesian policy with whole-body constraints through repeated candidate denoising. On PlaceToast with Panda, under the same RTX 4090, batch-1, CUDA-synchronized conditions, this adaptation takes 8.65 s per chunk with 16 proposals and 16 denoising steps, versus 0.254 s for EmbodiSteer. Reducing JM2D to four steps still takes 3.27 s per chunk. This cost limits its practicality for closed-loop replanning in our setting, so the paper reports computational feasibility rather than a full task-performance comparison.

Method

Lift Cartesian denoising into joint space

Frozen Cartesian policy

The learned denoiser remains in the embodiment-agnostic end-effector action space.

Joint-space sampling

Forward kinematics and damped Jacobian updates lift the reverse-diffusion sample into the target robot's joints.

Whole-body guidance

CBF-inspired QP guidance steers robot-body motion away from known obstacles while preserving end-effector behavior.

Citation

BibTeX

@inproceedings{wang2026embodisteer,
  title={EmbodiSteer: Steering Embodiment-Agnostic Visuomotor Policies with Joint-Space Guidance for Zero-Shot Cross-Embodiment Deployment},
  author={Shihefeng Wang and Kangchen Lv and Mingrui Yu and Xiang Li},
  booktitle={10th Annual Conference on Robot Learning},
  year={2026},
  url={https://openreview.net/forum?id=2Z8xQxrlsY}
}