DyWA: Dynamics-adaptive World Action Model for Generalizable Non-prehensile Manipulation
1. Paper overview
In one sentence: DyWA uses history-conditioned joint action and task-state prediction to improve single-view non-prehensile manipulation, while relying on privileged simulation supervision and a small real-world evaluation. e2e3e4e5e7e13e14e16e17
| At a glance | What to know |
|---|---|
| Research problem | Source description Move an object to a stable target 6D pose through pushing or flipping when grasping is inconvenient. A single camera hides geometry, and unknown mass or friction changes contact outcomes. The authors argue that a memoryless distilled policy averages across these conditions and that action imitation alone supplies insufficient dynamics supervision. e2e3 |
| Core mechanism | |
| A key reported result | Single-view 6D rearrangement with unknown object state: Seen: 82.2; unseen: 75.0. Success rate (%). IsaacGym; 323 training assets; unseen evaluation uses 10 objects at five scales. Stable start/goal poses; success within 0.05 m and 0.1 radians. CORN (PN++): 50.7/49.4; original CORN: 29.0/29.8, in seen/unseen order. Against the matched-backbone CORN baseline, gains are 31.5 and 25.6 percentage points, respectively. The headline “31.5%” is not a universal minimum gain across table columns. e11e12e13 |
| Reading caution | Author claim Point-cloud-only perception struggles with symmetric geometry and incomplete depth from transparent or specular objects. The authors propose adding appearance information. e20 |
Core contributions
Figure 2. History conditions a single student that predicts an action and its next task state. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the student on the left: current partial point cloud, end-effector pose and joint state enter modality encoders, while the initial cloud is transformed into a goal cloud. The upper branch compresses history, then decodes an embedding into FiLM conditioning. Follow the two outputs of the blue world action block: next task state and student action. On the right, the privileged teacher supplies action and adaptation targets through dashed orange links. The fire and snowflake distinguish the trainable student from the frozen teacher during distillation. The caption and Section 3.3 identify joint prediction; there is no explicit current-action input to a separate forward simulator. e3e4e5e7e8e21
What it supports. The central mechanism combines a shared action/state predictor with a separate conditioning path inside the student policy. Teacher supervision makes the history embedding approximate information available only in simulation. This supports the main One Model × Joint prediction classification, but the predicted state is an object-to-goal transformation rather than a video.
Where the evidence stops. History indices conflict: Figure 2 uses observations t−i+1 and actions t−i, whereas Equation (3) uses t−i−1 and t−i−2. Both pair observations with previous actions; the exact window endpoint is unresolved.
2. Motivation
2.1 The problem and the proposed response
Move an object to a stable target 6D pose through pushing or flipping when grasping is inconvenient. A single camera hides geometry, and unknown mass or friction changes contact outcomes. The authors argue that a memoryless distilled policy averages across these conditions and that action imitation alone supplies insufficient dynamics supervision. e2e3
2.2 What this reading follows
A push that works on one object can fail when the object is heavier, partly hidden, or resting on a different surface. DyWA addresses this uncertainty by combining two learning signals: historical interactions estimate missing geometry and physics, while next-state prediction gives action learning an explicit outcome target. The deployed student uses a partial point cloud and robot proprioception, with a goal expressed geometrically. Read the architecture first, then separate the evidence from simulation, component ablations, and physical trials. The key question is whether prediction and adaptation improve executed control; the paper does not establish a general simulator for testing arbitrary actions. e2e3e4e5e7e13e14e16e17
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | One Model |
| Prediction paradigm | Joint prediction |
| Quadrant | Q1 · One Model × Joint prediction |
3.1 Evidence-based assessment
Conflicts with the recorded classification
The principal WAM / One Model × Joint prediction quadrant is supported: Figure 2 shows one deployed student jointly producing an action and next task state, with history conditioning and a training-only teacher. It is not inverse dynamics. However, the recorded subcategories “Joint video-action modeling” and “3D multiview modeling” misdescribe the principal method: prediction is a low-dimensional task pose, and deployment uses one partial point-cloud view. Multiview appears as a benchmark condition. The conflict concerns these subcategories; the recorded snapshot is preserved. e4e5e11
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Build a goal without tracking the current object pose
Begin with the task representation, because it determines what the world model must predict. The paper defines the goal as a transformation from the initial object pose, and the task state as the transformation between the current object and that goal. In simulation, applying the goal transformation to the initial partial cloud creates a visual target. The student encodes that target together with the current partial cloud and robot proprioception; it need not receive a tracked current object pose. The next-state target remains available during training because simulation supplies the outcome after action execution. Physical evaluation obtains the goal cloud differently: the experimenter records the object at the desired pose before resetting the start. This is a concrete goal-acquisition requirement, even though online pose tracking is absent. e3e5e6e16
5.2 Use future-state prediction to shape a history-conditioned policy
The student must learn both what to do and what its action will change. Action imitation uses the privileged teacher as a target; next-state supervision uses the simulator outcome. A separate adaptation loss teaches the history encoder to approximate teacher geometry and physics embeddings. Those inferred features are decoded and injected through FiLM, so they can influence the same observation differently under different contact conditions. Reader interpretation: this design can make dynamics-relevant information useful to action learning without introducing an explicit planning loop. The state head predicts the outcome associated with the policy’s action, and Section 3.3 does not provide arbitrary candidate actions as inputs. After action prediction, inverse kinematics and variable impedance control execute the command. The architecture therefore supports joint policy/state prediction, while leaving the benefit of imagined multistep search untested. e4e5e6e7e8e9e10
5.3 Ask which comparison actually establishes generalization
The simulation benchmark separates information availability from object novelty. Its most useful headline comparison is the unknown-state, single-view track against CORN with the same point-cloud backbone: this reduces one architecture mismatch while retaining the difficult sensing condition. The component table then asks whether world prediction and adaptation conditioning matter inside DyWA. Reader interpretation: their joint benefit is more persuasive when the unseen column is considered alongside the seen column, because adding both modules without FiLM improves the latter much more than the former. Physical trials test a different boundary: transferring the simulated policy to a real camera, robot and contact process. Their results are encouraging but use few attempts, symmetry-aware scoring and a different baseline tracking pipeline. The friction table further narrows the mechanism question to one object across four surfaces. e11e12e13e14e16e17e18
5.4 Training and inference
During training
Train a privileged teacher with PPO for 200K simulation iterations, then a student with DAgger for 500K iterations. Figure 2 marks the teacher as frozen during distillation. Teacher actions supervise imitation; simulator outcomes supervise next state; teacher geometry/physics embeddings supervise adaptation. e4e6e7e10
Randomize object mass, scale and friction, and restitution of object, table and gripper. Add perturbations to torque commands, point clouds and goal poses during student training. Exact randomization ranges and perturbation magnitudes are absent from the supplied main paper. e10
During inference
Deploy the student without real-world fine-tuning or an external object-pose tracker. Recompute actions from observations and history in closed loop. The reported pipeline does not use predicted next states for candidate-action search or multistep rollout planning. e4e5e9e16
In physical evaluation the goal cloud is recorded by first placing the object at the target, then moving it to a random stable start. ICP measures final error afterward; it is evaluation machinery, not the student’s online pose-tracking input. e16
5.5 Implementation flow
- Encode geometry, robot state and goal
A simplified PointNet++ encodes current and goal point clouds with shared weights. Separate shallow MLPs encode joint state and end-effector pose. The goal representation avoids requiring the current object pose as a policy input. e3e5
- Infer a dynamics context
A 1D CNN processes a length-L history of observation and previous-action embeddings. Its adaptation embedding is supervised by concatenated teacher embeddings of full geometry and physical parameters, then decoded into a dynamics embedding. The exact history indexing is inconsistent across the figure and equation. e7e21
- Modulate and jointly predict
Dynamics-conditioned shallow MLPs produce FiLM scale and shift parameters for early student layers; final layers remain unconditioned. The shared world action model outputs both the action and next task state. The current action is not an explicit world-model input; the authors describe conditioning on the policy action as implicit. e4e5e8
- Execute through impedance control
The policy predicts an end-effector residual plus seven positional gains and seven damping factors. Damped-least-squares inverse kinematics yields a desired joint position, then joint impedance control executes the command through Polymetis. This controller, rather than the predicted task state itself, acts on the robot. e9
6. Experiments & results
DyWA learns contact-rich object rearrangement from a single-view partial point cloud. A privileged simulation teacher supervises a student that jointly predicts robot actions and the next object-to-goal transformation. Historical observations and actions supply a dynamics embedding that modulates the student through FiLM. The strongest evidence combines controlled simulation ablations with a small physical-robot evaluation; the method does not demonstrate planning through imagined rollouts.
6.1 Read the original evidence
Table 1. The largest matched-backbone gains occur with unknown object state and one camera. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the three column groups as different information conditions, not successive training stages. Each has a seen/unseen pair. The rightmost pair is the intended difficult deployment setting: no known current object pose and one camera. Compare DyWA with CORN (PN++), which receives the same vision backbone, before comparing with original CORN. The final Baselines sentence on page 6 specifies that all methods receive the same goal cloud when the current object pose is unknown. HACMan’s parenthesized values use a different, planar 3DoF goal according to the caption; do not pool them with full 6D rearrangement. Section 4.1 defines success as final errors within 0.05 m and 0.1 radians, with dynamics randomized in training and evaluation. e11e12e13
What it supports. DyWA scores 82.2%/75.0% on seen/unseen objects, compared with CORN (PN++) at 50.7%/49.4%. Subtracting within each column gives gains of 31.5 and 25.6 percentage points. With known state and three views, the corresponding rows are much closer, supporting the importance of the observation setting.
Where the evidence stops. The paper’s broad “at least 31.5%” wording is not valid as a minimum across these columns. Unseen geometry and randomized testing do not establish extrapolation beyond the training dynamics ranges, which are unspecified here.
Table 3. Physical trials support transfer across several object conditions, with limited per-object sampling. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each object column reports successes out of five attempts, not a percentage. The rightmost total aggregates ten objects. Compare the methods within each object, then inspect the slippery YCB bottle and the two non-uniform-mass columns. The real-world protocol uses a Franka arm and a side-view RealSense D435. Before execution, a goal cloud is recorded by placing the object at the target, after which the object is moved to a random stable start. Final ICP alignment measures error. For symmetric objects, evaluation ignores ambiguous symmetry-axis components, so the protocol does not demand unique orientation recovery in every case. e16e17
What it supports. DyWA achieves 34/50 successful trials, or 68%, against CORN with tracking at 18/50, or 36%. The half-full bottle changes from 0/5 to 4/5. The card and book tie at 4/5 each, and DyWA’s chips-can result remains 2/5, showing that the aggregate improvement does not imply uniform reliability.
Where the evidence stops. Five attempts per object provide limited precision. CORN uses a tracker while DyWA does not, so this comparison evaluates whole sensing-and-control pipelines. Symmetry-relaxed scoring also narrows the meaning of 6D success.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Single-view 6D rearrangement with unknown object state IsaacGym; 323 training assets; unseen evaluation uses 10 objects at five scales. Stable start/goal poses; success within 0.05 m and 0.1 radians. | Seen: 82.2; unseen: 75.0. Success rate (%) | CORN (PN++): 50.7/49.4; original CORN: 29.0/29.8, in seen/unseen order. Against the matched-backbone CORN baseline, gains are 31.5 and 25.6 percentage points, respectively. The headline “31.5%” is not a universal minimum gain across table columns. e11e12e13 |
| Contribution of world prediction and FiLM Table 2; unknown-state, single-view simulation; seen/unseen objects. | Full model: 82.2/75.0. Success rate (%) | Without world modeling: 70.0/63.7; without FiLM: 73.3/59.4; DAgger: 59.9/57.5. Both auxiliary prediction and the conditioning mechanism matter in this configuration. The table supplies no uncertainty estimates or seed counts. e14 |
| Zero-shot real-world object rearrangement Franka arm, side-view RealSense D435; 10 unseen objects, five attempts each. Recorded goal clouds and final ICP evaluation; symmetry axes are relaxed. | 34/50 (68%). Successful trials / attempts (%) | CORN with tracking: 18/50 (36%); half-full bottle: DyWA 4/5 versus CORN 0/5. Physical execution improves by 32 percentage points on this small object set. Per-object conclusions rest on only five attempts, and sensing pipelines differ. e16e17 |
| Adaptation across four table-friction conditions Physical bulldozer toy; four ordered-friction tablecloths, five attempts per condition. | Full model: 4/5 at every level; times 45, 50, 49, 51 s. Successful trials and mean time over successful episodes | Without adaptation: 3/5, 3/5, 4/5, 3/5; times 65, 81, 96, 124 s. Adaptation accompanies steadier successful-run times. Exact friction coefficients and failure durations are not reported. e18 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. The complete combination improves both splits; isolated modules provide smaller gains. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read W.M. as world modeling and D.A. as dynamics adaptation; checks indicate included components. All rows use the unknown-state, single-view track. Start with DAgger, then examine the world-model-only and RMA rows. Next compare the full model with “Ours w/o W.M.” to retain adaptation and FiLM while removing future-state supervision. Compare it with “Ours w/o FiLM” to assess the conditioning mechanism when both prediction and adaptation remain. Keep the unseen column visible: improvements on seen objects alone can give a misleading impression of how well a component combination transfers to unfamiliar shapes. e11e14
What it supports. The full model reaches 82.2%/75.0%, versus 70.0%/63.7% without world modeling and 73.3%/59.4% without FiLM. The combination without FiLM improves seen-object success over DAgger, but its unseen score is only 59.4% versus 57.5%. This makes the conditioning design especially relevant to the reported unseen-object result.
Where the evidence stops. These rows support component usefulness in this setup, but do not fully separate representation, parameter-count and rollout-distribution effects. The table reports neither variability across seeds nor confidence intervals.
Figure 3. Joint prediction and adaptation improve the reported training-loss trajectories. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the panels separately because they use different loss scales. The left panel compares imitation loss for dynamics adaptation alone with world modeling plus adaptation. The right compares world-model loss without adaptation against the same combined model. In both panels the orange curve denotes the combination, and the horizontal axis is iterations multiplied by 100,000. The caption attributes the unusual early transient to DAgger beginning with teacher action execution and progressively introducing student actions. Focus on later trends and the accompanying success ablation, rather than interpreting the initial dip as immediate mastery of the manipulation task. e10e14e15
What it supports. The combined model ends with lower imitation loss than adaptation alone and lower world-model loss than prediction without adaptation. This is consistent with the authors’ proposed complementarity: history helps predict outcomes, and outcome supervision helps learn actions. Table 2 is needed to connect those optimization trends to task success.
Where the evidence stops. The curves show training diagnostics without uncertainty bands. They do not establish calibrated prediction errors on held-out dynamics, and the changing DAgger action distribution complicates a purely representation-based causal interpretation.
Table 4. History-based adaptation accompanies steadier successful-run times across four surfaces. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each friction group contains two measures: success rate, where larger is better, and average time, where smaller is better. The caption orders the tablecloth conditions as μ1 < μ2 < μ3 < μ4; it does not assign numerical coefficients. The physical test uses one bulldozer toy on all four surfaces. Read success counts before the colored times: a short completion time says nothing about failed trials. Section 4.3 explicitly restricts the average to successful episodes. Compare the full model against the no-adaptation row within each surface to keep the physical condition aligned. e18
What it supports. The full model succeeds in 4/5 attempts at every friction level, with mean successful-run times of 45, 50, 49 and 51 seconds. Without adaptation, success is 3/5, 3/5, 4/5 and 3/5, and times increase from 65 to 124 seconds across the ordered surfaces. These observations support a targeted adaptation benefit.
Where the evidence stops. This is one object and four surfaces, with five attempts per condition. Missing friction measurements, failure durations and uncertainty prevent a general quantitative claim about speed or robustness as a function of friction.
7. Analysis & limitations
7.1 What the evidence leaves open
Point-cloud-only perception struggles with symmetric geometry and incomplete depth from transparent or specular objects. The authors propose adding appearance information. e20
Simulation randomizes dynamics in both training and testing, so unseen geometry does not by itself establish extrapolation beyond the training physics range. Physical success criteria relax symmetry, limiting comparison with strict full-orientation accuracy. e11e16
Figure 3 shows optimization trends, not calibrated long-horizon dynamics accuracy. Language handover and pre-grasping examples demonstrate integration, but the supplied paper gives no quantitative application benchmark. e15e19
7.2 Questions for discussion
- Does a correctly aligned history improve physical-parameter identification, or mostly compensate for occluded geometry?
- Does next-state supervision still improve action success under a held-out dynamics distribution and matched rollout data?
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction needs the IsaacGym assets/splits, CORN-style teacher reward, student architecture and history length, DAgger action-mixing schedule, randomization ranges, and controller settings. The main paper gives training iteration counts and success thresholds but leaves these implementation details incomplete; it also omits training hardware, runtime and software versions. e3e5e7e9e10e11
Resolve the history alignment before implementing adaptation: Figure 2 labels observations t−i+1 and actions t−i, while Equation (3) prints t−i−1 and t−i−2. Section 3.4 consistently pairs an observation with its previous action, but the window endpoint remains unresolved. e21
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test whether accurate outcome supervision helps actions
Reader-proposed check, not a completed experiment: compare three otherwise matched students with adaptation and FiLM enabled—correct next-state labels, no world loss, and shuffled next-state labels. Hold the teacher checkpoint, backbone, action loss, randomization draws and training budget fixed. First use an identical recorded transition stream to control data exposure, then repeat with the same DAgger mixing schedule to evaluate executed control. Measure held-out next-state error and seen/unseen success under the paper’s pose thresholds across multiple seeds. The mechanism is supported if correct labels improve success beyond both controls; an equal gain from shuffled labels would weaken the claim that learning meaningful outcomes is responsible. e4e6e10e11e14e15
Check 2: Resolve the history window and isolate temporal adaptation
Reader-proposed check, not a completed experiment: implement the recent history window shown in Figure 2 and the delayed window printed in Equation (3), keeping each observation paired with its preceding action. Add a control that permutes the order of those same pairs. Keep history length, network size, teacher, controller and physics draws fixed; explicitly document the selected history length because the main paper omits it. Evaluate both stable and deliberately changed friction conditions, recording next-state error, success, time to recover after the change, and durations of failures as well as successes. A consistent advantage for ordered recent history would support temporal adaptation; comparable permuted-history performance would favor a weaker explanation based on pooled geometry or average dynamics. e7e9e11e18e21
8.3 Reading coverage
Visual audit: Visually inspected the title and author block, CVF edition statement and Figure 1 on page 1; motivation on page 2; Figure 2 and task/teacher formulation on page 3; joint prediction, losses, history indices and FiLM on page 4; controller, training, benchmark protocol and Figure 3 on page 5; Tables 1–2 and ablation text on page 6; Tables 3–4, physical setup, scoring and friction protocol on page 7; Figures 4–6, applications and limitations on page 8. All six final original crops were individually viewed at their returned resolutions. Figure 2 history indices were cross-checked against Equation (3), with their disagreement disclosed. Equations (2) and (4) were checked visually for squared norms. Page 6 was additionally inspected at 300 DPI: the final Baselines sentence before Results, in the left column, specifies unknown current object pose for the shared goal-cloud representation. Table explanations retain the planar-goal caveat, symmetry scoring and successful-episode timing restriction. References on pages 9–11 were read in text; no separate supplementary PDF was supplied.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract (PDF p. 1)
- 1. Introduction (pp. 1–2)
- 2. Related Works, including 2.1–2.2 (pp. 2–3)
- 3. Method, including 3.1–3.6 (pp. 3–5)
- 4. Experiments, including 4.1–4.4 (pp. 5–8)
- 5. Conclusion, Limitations, and Future Works (p. 8)
- Acknowledgements (p. 8)
- References (pp. 9–11)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- All five supplied text chunks were read. PDF pages 1–8 and all six final crops were visually inspected; references on pages 9–11 were read as text.
- No appendix is present in this 11-page PDF. Separately referenced supplementary material was not supplied, including reward and architecture details, stable-pose preparation, additional conditioning experiments, teacher-world-model experiments, and setup illustrations.
- The title page identifies the CVF Open Access edition as identical to the accepted version except for its watermark and distinguishes the final IEEE proceedings version. Title and all six authors match the catalog. No separate revision identifier/date is stated; no other edition was inspected.
- Code, project pages, linked works and external supplements were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e1PDF p. 1 (printed p. 11058), title/author block and CVF watermark
Title and author order match the catalog: Jiangran Lyu, Ziming Li, Xuesong Shi, Chaoyi Xu, Yizhou Wang, He Wang. The watermark identifies the CVF Open Access accepted-paper edition and distinguishes the final IEEE proceedings version.
Go to primary source ↓e2PDF pp. 1–2, Abstract and Section 1
The problem is non-prehensile rearrangement under partial geometry and varying physical conditions; the authors motivate history-based adaptation and future-state supervision.
Go to primary source ↓e3PDF p. 3, Sections 3.1–3.2
G is the goal relative to the initial object pose; S_t is the current-to-goal transformation. Observations comprise partial point cloud, joint state and end-effector pose. A privileged teacher has full geometry, task state and physical parameters; reward details are deferred to the supplement.
Go to primary source ↓e4PDF p. 3, Figure 2 and caption
The trainable student has a shared world action block outputting S_{t+1} and A_t^s. History passes through adaptation and a decoder into FiLM. Dashed supervision links connect a teacher marked with a snowflake to student action and adaptation targets.
Go to primary source ↓e5PDF p. 4, Section 3.3, Definition and Observation and Goal Encoding; State-based World Modeling
A simplified PointNet++ shares observation/goal encoding; shallow MLPs encode proprioception. P_G=GP_0. MLP processing jointly predicts action and next task state; current action is not explicitly supplied as input. Architecture details are deferred to the supplement.
Go to primary source ↓e6PDF p. 4, Section 3.3, Equations (1)–(2)
World loss is squared translation L2 plus rotation L1; un-hatted transformations are predictions and hatted ones are simulation outcomes. A 9D rotation representation is used. The printed imitation equation uses a squared student–teacher action norm.
Go to primary source ↓e7PDF p. 4, Section 3.4, Adaptation Embedding and Equations (3)–(4)
A length-L observation/action history enters a 1D CNN. Adaptation is supervised with teacher full-geometry and physics embeddings using a printed squared norm. The paper does not specify L numerically here.
Go to primary source ↓e8PDF pp. 4–5, Section 3.4, Dynamics Conditioning and Equation (5)
A decoder produces dynamics conditioning; two shallow MLPs produce FiLM scale gamma and shift beta. Modulation is dense in early world-action layers, while final layers are unconditioned.
Go to primary source ↓e9PDF p. 5, Section 3.5 and Equation (6)
Actions contain an end-effector residual and seven-dimensional positional gains P and damping factors rho; D=rho sqrt(P). Damped-least-squares IK produces q_d=q_t+IK(delta T_ee); joint impedance executes through Polymetis. The final prose switches the positional-gain symbol to K without defining a separate quantity.
Go to primary source ↓e10PDF p. 5, Section 3.6 and Equation (7)
Teacher PPO runs for 200K iterations, student DAgger for 500K. The three losses are summed. Domain randomization varies mass, scale, friction and restitution; student perturbations affect torque commands, point clouds and goals. Numerical ranges, noise magnitudes, training hardware and runtime are not given.
Go to primary source ↓e11PDF p. 5, Section 4.1, benchmark and Task Setup paragraphs
IsaacGym benchmark uses 323 DexGraspNet assets for training and 10 unseen geometries at five scales. Both training and testing randomize dynamics. Camera/state tracks vary observation access. Stable goals are at least 0.1 m from starts; success requires 0.05 m and 0.1 radian errors. Stable-pose preparation is deferred to the supplement.
Go to primary source ↓e12PDF pp. 5–6, Section 4.1, Baselines; p. 6 (printed p. 11063), left column, final Baselines sentence immediately before Results; Table 1 caption
HACMan is reimplemented in IsaacGym and its point-correspondence requirement causes difficulty without state. CORN PN++ uses the DyWA vision backbone. The final Baselines sentence specifies the condition as ‘where the current object pose is unknown’ when stating that all methods receive the same goal point cloud representation. HACMan parenthesized numbers use a separate planar 3DoF goal.
Go to primary source ↓e13PDF p. 6, Table 1, all tracks and Ours/CORN (PN++)/CORN rows; Results paragraph
Unknown-state one-view success is 82.2/75.0 for DyWA, 50.7/49.4 for CORN PN++, and 29.0/29.8 for CORN (seen/unseen). DyWA has 87.9/85.0 with known state/three views and 85.8/82.3 with unknown state/three views. The prose headline says at least 31.5% improvement, which is not the difference in every comparison.
Go to primary source ↓e14PDF p. 6, Table 2 and Section 4.2; p. 7, Effectiveness of FiLM Conditioning
Seen/unseen success: DAgger 59.9/57.5; world model only 61.6/59.4; RMA 65.6/57.9; no world model with adaptation+FiLM 70.0/63.7; no FiLM with world model+adaptation 73.3/59.4; full 82.2/75.0. FiLM is compared with input concatenation in the accompanying text. No seed counts or uncertainty intervals are supplied.
Go to primary source ↓e15PDF p. 5, Figure 3 and caption; p. 6, Section 4.2, Synergy paragraph
The plots compare imitation loss with/without world modeling under dynamics adaptation, and world-model loss with/without adaptation. The combined variant finishes lower in both panels. The caption explains an early teacher-to-student action-mixing transition under DAgger; no uncertainty bands are shown.
Go to primary source ↓e16PDF p. 7, Section 4.3, Real-World Setup and Generalization across Diverse Objects
A Franka arm and side-view RealSense D435 evaluate 10 unseen objects. A target cloud is recorded before each randomized start. Final ICP error evaluates success, with symmetric axes relaxed. DyWA does not use external pose tracking; the CORN comparator does.
Go to primary source ↓e17PDF p. 7, Table 3, all rows and Avg./Half-full Bottle columns
DyWA achieves 34/50 (68%) versus CORN with tracking 18/50 (36%). Each object has five attempts. Half-full bottle success is 4/5 versus 0/5; card and book are ties at 4/5 each; DyWA chips-can success is 2/5.
Go to primary source ↓e18PDF p. 7, Table 4 and Section 4.3, Robustness to Surface Friction Variations
The bulldozer toy is evaluated on four tablecloths ordered mu_1<mu_2<mu_3<mu_4. Full model has 4/5 success at each level and successful-run means of 45/50/49/51 seconds; no-adaptation has 3/5,3/5,4/5,3/5 and 65/81/96/124 seconds. Numerical friction coefficients and failure times are not given.
Go to primary source ↓e19PDF pp. 7–8, Section 4.4; p. 8, Figures 4–6 and captions
SoFar supplies semantic goal poses for a language-guided drill handover. Qualitative sequences show rearrangement and pre-grasping a thin card and broad cracker box. No quantitative language-handover or pre-grasping evaluation is supplied.
Go to primary source ↓e20PDF p. 8, Section 5, Conclusion, Limitations, and Future Works
Authors identify geometric ambiguity for symmetric objects and incomplete depth for transparent/specular objects as limitations of point-cloud-only input, suggesting appearance information as future work.
Go to primary source ↓e21PDF p. 3, Figure 2, upper-left history labels; p. 4, Section 3.4 and Equation (3)
Figure 2 prints observation features at t−i+1 and actions at t−i, for i=1 through L; Equation (3) prints observation features t−i−1 and action features t−i−2. The prose pairs a current observation embedding with its previous action embedding. These descriptions disagree on the history window endpoint.
Go to primary source ↓8.5 Primary sources
DyWA: Dynamics-adaptive World Action Model for Generalizable Non-prehensile Manipulation ↗
PDF · 7,403 extracted words
Source fingerprint
056f1e5038cd0bf03eb705e731be260d99f7e49dfc2495b69cf0c390ce870868